kfallah commited on Aug 1

Commit

ecc1989

verified ·

1 Parent(s): 408a34b

Trained default: EXP-012 reward_lcb_b0.2 (tuned fp16 + tuned 4-bit MLX)

Browse files

Files changed (18) hide show

.gitattributes +2 -0
README.md +77 -27
encoder-fp16/1_Pooling/config.json +10 -0
encoder-fp16/chat_template.jinja +85 -0
encoder-fp16/config.json +63 -0
encoder-fp16/config_sentence_transformers.json +8 -0
encoder-fp16/model.safetensors +3 -0
encoder-fp16/modules.json +20 -0
encoder-fp16/tokenizer.json +3 -0
encoder-fp16/tokenizer_config.json +15 -0
encoder-mlx-4bit/chat_template.jinja +85 -0
encoder-mlx-4bit/config.json +71 -0
encoder-mlx-4bit/model.safetensors +3 -0
encoder-mlx-4bit/model.safetensors.index.json +711 -0
encoder-mlx-4bit/tokenizer.json +3 -0
encoder-mlx-4bit/tokenizer_config.json +15 -0
router.json +578 -571
router.npz +2 -2

.gitattributes CHANGED Viewed

@@ -33,3 +33,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text

 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+encoder-fp16/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+encoder-mlx-4bit/tokenizer.json filter=lfs diff=lfs merge=lfs -text

README.md CHANGED Viewed

@@ -1,39 +1,89 @@
 ---
 tags:
-- routing
-- code
-library_name: coding-router
 ---
-# coding-router — default routing artifact
-The default kNN routing artifact for [coding-router](https://github.com/experientiallabs/coding-router):
-routes every coding-agent request to the cheapest OpenAI/Anthropic model likely to solve it.
-- `router.json` — arm specs (model + reasoning effort + pricing), kNN hyperparameters
-  (k=12, tau=0.5, sim_floor=0.35), fallback arm, provenance.
-- `router.npz` — reference-task embeddings plus per-arm resolved/cost cells measured on
-  DeepSWE tasks across 41 OpenAI/Anthropic arms.
-Embedding space: **Qwen3-Embedding-0.6B**, computed fully locally at serve time
-(`mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ` via MLX on Apple Silicon;
-`Qwen/Qwen3-Embedding-0.6B` via sentence-transformers elsewhere). Routing decisions never
-call an embeddings API. Off-distribution queries abstain to the strongest arm rather than
-trusting a thin neighborhood.
-Holdout (repo-grouped 80/20 over DeepSWE): quality parity with the always-strongest-arm
-baseline at a fraction of its cost (median cost ratio ≈ 3.8× cheaper across seeds in this
-embedding space).
-## Use
-You never fetch this by hand — `uv run python -m router.serve` downloads it on first run,
-then runs offline. New router versions (including future RL-trained ones) overwrite this
-repo in place; old versions stay in the repo's git history.
-Fit your own artifact from your own traces with
-[world-model-optimizer](https://github.com/experientiallabs/world-model-optimizer)
-(`wmo optimize route`).
-This artifact contains only benchmark-task embeddings and measured outcomes — no user
-data of any kind.

 ---
+license: apache-2.0
+base_model: Qwen/Qwen3-Embedding-0.6B
 tags:
+  - router
+  - model-routing
+  - embeddings
+  - mlx
 ---
+# coding-router — trained default artifact (v1)
+The default routing artifact for [experiential-labs/coding-router](https://github.com/experiential-labs/coding-router):
+given a coding task (or a whole agent conversation), pick the cheapest OpenAI/Anthropic
+model+effort arm predicted to solve it. Routing runs fully locally — the encoder below
+runs in-process; API keys are only ever used to dispatch the chosen model.
+This release replaces the kNN-over-base-embeddings artifact with a **trained** router
+(EXP-012 winning recipe `reward_lcb_b0.2`), shipped in two encoder formats plus one
+head/calibration payload.
+## Files
+| path | what |
+|---|---|
+| `router.json` | arm list + per-arm request kwargs, decision-rule params (`T`, `lam`, `sim_floor`), provenance (`kind: "trained"`) |
+| `router.npz` | `emb` — 110-task DeepSWE evidence bank in the tuned space; `graded` — 41×110 outcome matrix (the calibration side of the vote); `med_cost` — per-arm median $/task |
+| `encoder-fp16/` | merged (LoRA→base) Qwen3-Embedding-0.6B, fp16 safetensors — loads via sentence-transformers / transformers on CUDA or CPU |
+| `encoder-mlx-4bit/` | the same merged encoder, 4-bit MLX (group size 64) for Apple Silicon — loads via `mlx_embeddings` |
+Both encoder dirs are the SAME model in two precisions; the bank and every query must
+be embedded by one of them (backend agreement measured below). The server picks MLX on
+Apple Silicon and torch elsewhere, and downloads only the directory it needs.
+## What the router is
+Qwen3-Embedding-0.6B with LoRA r=16 on the attention projections, trained by exact
+expected reward −E_π[graded − λ·cost] over the LiveCodeBench 7×76 outcome matrix with a
+KL(π‖π_init) anchor (β=0.2), plus a trained soft-vote temperature `T`. At inference:
+embed the task → softmax(sims/`T`) vote over the DeepSWE bank → per-arm P(solve) →
+argmax of u = P(solve) − `lam`·med_cost. If no bank task is within `sim_floor` cosine
+similarity the router abstains: it escalates to the strongest arm rather than trust the
+vote off-distribution.
+Final-mint policy: one training run with the sweep's exact code and seed; (checkpoint,
+`lam`) selected only on the six standard seeds' inner DeepSWE-train splits (the sweep's
+own selection discipline); the shipped bank is the full 110-task evidence. The artifact
+was never selected or early-stopped on the holdout numbers reported below.
+## Honest numbers (EXP-012, 6-seed repo-split holdout, DeepSWE v1.1)
+Two baselines, deliberately:
+| policy | graded | cost/split | read |
+|---|---|---|---|
+| always-best-train arm (deployable baseline) | 0.9336 | $126.28 | what you could actually deploy without hindsight |
+| **this router (per-seed selection, sweep)** | **0.9484** | **$60.17** | **+0.015 graded, 2.1× cheaper** |
+| hindsight-best static arm (opus-5@high) | 0.9635 | $135.73 | the router is at parity with it, NOT better |
+Minted-artifact re-evaluation (this exact shipped artifact's encoder + single global
+(step=50, lam=0.01) instead of per-seed selection, same 6 splits, sweep eval code):
+graded **0.9353 at $50.21/split** — slightly below the per-seed-selected sweep mean
+(that gap is the per-seed selection's optimism, deliberately not shipped), cheaper,
+and still above the always-best-train baseline on both axes. The shipped checkpoint
+and `lam` were selected on inner train-side splits only; the holdout numbers above
+were computed after the artifact was frozen.
+Quality-BEATING routing on this data is dead (held-out oracle analysis, EXP-018): the
+honest framing is parity-quality at a large cost saving vs any deployable policy.
+**Cost figures are matrix-based (June 2026 collection) and pending live re-benchmark
+(EXP-014 measured live cost/quality drift on several arms).**
+## Backend agreement (fp16 vs MLX 4-bit)
+Over 20 probe texts (DeepSWE issues, LCB tasks, short interactive prompts): max
+per-arm |ΔP(solve)| = **0.0274**; decisions and abstentions agreed on **20/20**
+probes, no arm-ranking flips (`provenance/agreement.json`). Cross-backend embedding
+cosine runs 0.91–0.97 — the temperature-softmax vote absorbs the quantization noise.
+## Scope warning
+Validated on repo-issue-shaped inputs (p50 ≈ 2,000 chars, long-horizon SWE tasks).
+Short one-line prompts sit below `sim_floor` and abstain to the strongest arm — safe,
+but no saving. Always check `Decision.off_distribution`.
+## Rollback
+The previous kNN artifact keeps loading with the same product code (`kind` defaults to
+`"knn"`); pin the prior revision of this repo to roll back.

encoder-fp16/1_Pooling/config.json ADDED Viewed

	@@ -0,0 +1,10 @@

+{
+    "word_embedding_dimension": 1024,
+    "pooling_mode_cls_token": false,
+    "pooling_mode_mean_tokens": false,
+    "pooling_mode_max_tokens": false,
+    "pooling_mode_mean_sqrt_len_tokens": false,
+    "pooling_mode_weightedmean_tokens": false,
+    "pooling_mode_lasttoken": true,
+    "include_prompt": true
+}

encoder-fp16/chat_template.jinja ADDED Viewed

	@@ -0,0 +1,85 @@

+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

encoder-fp16/config.json ADDED Viewed

	@@ -0,0 +1,63 @@

+{
+  "architectures": [
+    "Qwen3Model"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float16",
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.14.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151669
+}

encoder-fp16/config_sentence_transformers.json ADDED Viewed

	@@ -0,0 +1,8 @@

+{
+  "prompts": {
+    "query": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:",
+    "document": ""
+  },
+  "default_prompt_name": null,
+  "similarity_fn_name": "cosine"
+}

encoder-fp16/model.safetensors ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:734f269be9b14ca9d861ff6f3b826dee8f73997c71bac68187e1bd22aa6b249c
+size 1191586112

encoder-fp16/modules.json ADDED Viewed

	@@ -0,0 +1,20 @@

+[
+  {
+    "idx": 0,
+    "name": "0",
+    "path": "",
+    "type": "sentence_transformers.models.Transformer"
+  },
+  {
+    "idx": 1,
+    "name": "1",
+    "path": "1_Pooling",
+    "type": "sentence_transformers.models.Pooling"
+  },
+  {
+    "idx": 2,
+    "name": "2",
+    "path": "2_Normalize",
+    "type": "sentence_transformers.models.Normalize"
+  }
+]

encoder-fp16/tokenizer.json ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:93623af029cdc69b87f2864d3b2cc2424fdf16684f15e139b5b9d08ec34ced91
+size 11423701

encoder-fp16/tokenizer_config.json ADDED Viewed

	@@ -0,0 +1,15 @@

+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

encoder-mlx-4bit/chat_template.jinja ADDED Viewed

	@@ -0,0 +1,85 @@

+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

encoder-mlx-4bit/config.json ADDED Viewed

	@@ -0,0 +1,71 @@

+{
+  "architectures": [
+    "Qwen3Model"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "float16",
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 1024,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_types": [
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 28,
+  "model_type": "qwen3",
+  "num_attention_heads": 16,
+  "num_hidden_layers": 28,
+  "num_key_value_heads": 8,
+  "pad_token_id": null,
+  "rms_norm_eps": 1e-06,
+  "rope_parameters": {
+    "rope_theta": 1000000,
+    "rope_type": "default"
+  },
+  "sliding_window": null,
+  "tie_word_embeddings": true,
+  "transformers_version": "5.14.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151669,
+  "quantization": {
+    "group_size": 64,
+    "bits": 4
+  },
+  "quantization_config": {
+    "group_size": 64,
+    "bits": 4
+  }
+}

encoder-mlx-4bit/model.safetensors ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:9bc318baec5626eea24d59051c8f2054ef105b0b79cbd430d35e930f303a5568
+size 335296249

encoder-mlx-4bit/model.safetensors.index.json ADDED Viewed

	@@ -0,0 +1,711 @@

+{
+    "metadata": {
+        "total_size": 335218496
+    },
+    "weight_map": {
+        "model.embed_tokens.biases": "model.safetensors",
+        "model.embed_tokens.scales": "model.safetensors",
+        "model.embed_tokens.weight": "model.safetensors",
+        "model.layers.0.input_layernorm.weight": "model.safetensors",
+        "model.layers.0.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.0.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.0.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.0.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.0.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.0.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.0.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.0.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.0.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.0.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.0.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.1.input_layernorm.weight": "model.safetensors",
+        "model.layers.1.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.1.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.1.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.1.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.1.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.1.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.1.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.1.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.1.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.1.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.1.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.10.input_layernorm.weight": "model.safetensors",
+        "model.layers.10.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.10.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.10.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.10.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.10.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.10.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.10.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.10.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.10.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.10.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.10.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.11.input_layernorm.weight": "model.safetensors",
+        "model.layers.11.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.11.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.11.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.11.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.11.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.11.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.11.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.11.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.11.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.11.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.11.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.12.input_layernorm.weight": "model.safetensors",
+        "model.layers.12.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.12.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.12.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.12.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.12.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.12.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.12.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.12.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.12.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.12.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.12.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.13.input_layernorm.weight": "model.safetensors",
+        "model.layers.13.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.13.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.13.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.13.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.13.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.13.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.13.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.13.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.13.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.13.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.13.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.14.input_layernorm.weight": "model.safetensors",
+        "model.layers.14.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.14.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.14.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.14.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.14.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.14.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.14.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.14.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.14.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.14.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.14.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.15.input_layernorm.weight": "model.safetensors",
+        "model.layers.15.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.15.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.15.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.15.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.15.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.15.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.15.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.15.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.15.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.15.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.15.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.16.input_layernorm.weight": "model.safetensors",
+        "model.layers.16.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.16.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.16.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.16.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.16.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.16.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.16.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.16.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.16.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.16.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.16.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.17.input_layernorm.weight": "model.safetensors",
+        "model.layers.17.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.17.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.17.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.17.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.17.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.17.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.17.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.17.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.17.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.17.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.17.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.18.input_layernorm.weight": "model.safetensors",
+        "model.layers.18.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.18.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.18.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.18.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.18.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.18.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.18.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.18.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.18.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.18.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.18.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.19.input_layernorm.weight": "model.safetensors",
+        "model.layers.19.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.19.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.19.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.19.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.19.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.19.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.19.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.19.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.19.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.19.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.19.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.2.input_layernorm.weight": "model.safetensors",
+        "model.layers.2.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.2.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.2.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.2.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.2.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.2.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.2.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.2.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.2.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.2.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.2.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.20.input_layernorm.weight": "model.safetensors",
+        "model.layers.20.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.20.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.20.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.20.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.20.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.20.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.20.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.20.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.20.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.20.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.20.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.21.input_layernorm.weight": "model.safetensors",
+        "model.layers.21.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.21.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.21.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.21.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.21.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.21.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.21.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.21.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.21.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.21.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.21.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.22.input_layernorm.weight": "model.safetensors",
+        "model.layers.22.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.22.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.22.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.22.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.22.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.22.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.22.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.22.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.22.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.22.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.22.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.23.input_layernorm.weight": "model.safetensors",
+        "model.layers.23.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.23.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.23.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.23.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.23.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.23.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.23.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.23.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.23.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.23.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.23.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.24.input_layernorm.weight": "model.safetensors",
+        "model.layers.24.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.24.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.24.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.24.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.24.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.24.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.24.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.24.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.24.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.24.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.24.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.25.input_layernorm.weight": "model.safetensors",
+        "model.layers.25.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.25.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.25.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.25.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.25.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.25.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.25.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.25.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.25.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.25.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.25.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.26.input_layernorm.weight": "model.safetensors",
+        "model.layers.26.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.26.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.26.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.26.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.26.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.26.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.26.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.26.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.26.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.26.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.26.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.27.input_layernorm.weight": "model.safetensors",
+        "model.layers.27.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.27.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.27.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.27.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.27.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.27.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.27.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.27.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.27.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.27.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.27.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.3.input_layernorm.weight": "model.safetensors",
+        "model.layers.3.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.3.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.3.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.3.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.3.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.3.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.3.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.3.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.3.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.3.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.3.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.4.input_layernorm.weight": "model.safetensors",
+        "model.layers.4.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.4.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.4.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.4.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.4.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.4.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.4.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.4.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.4.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.4.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.4.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.5.input_layernorm.weight": "model.safetensors",
+        "model.layers.5.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.5.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.5.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.5.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.5.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.5.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.5.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.5.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.5.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.5.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.5.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.6.input_layernorm.weight": "model.safetensors",
+        "model.layers.6.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.6.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.6.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.6.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.6.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.6.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.6.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.6.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.6.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.6.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.6.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.7.input_layernorm.weight": "model.safetensors",
+        "model.layers.7.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.7.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.7.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.7.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.7.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.7.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.7.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.7.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.7.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.7.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.7.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.8.input_layernorm.weight": "model.safetensors",
+        "model.layers.8.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.8.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.8.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.8.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.8.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.8.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.8.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.8.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.8.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.8.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.8.self_attn.v_proj.weight": "model.safetensors",
+        "model.layers.9.input_layernorm.weight": "model.safetensors",
+        "model.layers.9.mlp.down_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.down_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.down_proj.weight": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.gate_proj.weight": "model.safetensors",
+        "model.layers.9.mlp.up_proj.biases": "model.safetensors",
+        "model.layers.9.mlp.up_proj.scales": "model.safetensors",
+        "model.layers.9.mlp.up_proj.weight": "model.safetensors",
+        "model.layers.9.post_attention_layernorm.weight": "model.safetensors",
+        "model.layers.9.self_attn.k_norm.weight": "model.safetensors",
+        "model.layers.9.self_attn.k_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.k_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.k_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.o_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.q_norm.weight": "model.safetensors",
+        "model.layers.9.self_attn.q_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.q_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.q_proj.weight": "model.safetensors",
+        "model.layers.9.self_attn.v_proj.biases": "model.safetensors",
+        "model.layers.9.self_attn.v_proj.scales": "model.safetensors",
+        "model.layers.9.self_attn.v_proj.weight": "model.safetensors",
+        "model.norm.weight": "model.safetensors"
+    }
+}

encoder-mlx-4bit/tokenizer.json ADDED Viewed

	@@ -0,0 +1,3 @@

+version https://git-lfs.github.com/spec/v1
+oid sha256:93623af029cdc69b87f2864d3b2cc2424fdf16684f15e139b5b9d08ec34ced91
+size 11423701

encoder-mlx-4bit/tokenizer_config.json ADDED Viewed

	@@ -0,0 +1,15 @@

+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "is_local": false,
+  "local_files_only": false,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}

router.json CHANGED Viewed

@@ -1,573 +1,580 @@
 {
-  "version": "v0",
-  "embed_model": "mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ",
-  "arms": [
-    "mini_swe_agent_claude_fable_5_high",
-    "mini_swe_agent_claude_fable_5_low",
-    "mini_swe_agent_claude_fable_5_max",
-    "mini_swe_agent_claude_fable_5_medium",
-    "mini_swe_agent_claude_fable_5_xhigh",
-    "mini_swe_agent_claude_opus_4_8_high",
-    "mini_swe_agent_claude_opus_4_8_low",
-    "mini_swe_agent_claude_opus_4_8_max",
-    "mini_swe_agent_claude_opus_4_8_medium",
-    "mini_swe_agent_claude_opus_4_8_xhigh",
-    "mini_swe_agent_claude_opus_5_high",
-    "mini_swe_agent_claude_opus_5_low",
-    "mini_swe_agent_claude_opus_5_max",
-    "mini_swe_agent_claude_opus_5_medium",
-    "mini_swe_agent_claude_opus_5_xhigh",
-    "mini_swe_agent_claude_sonnet_4_6_high",
-    "mini_swe_agent_claude_sonnet_5_high",
-    "mini_swe_agent_claude_sonnet_5_low",
-    "mini_swe_agent_claude_sonnet_5_max",
-    "mini_swe_agent_claude_sonnet_5_medium",
-    "mini_swe_agent_claude_sonnet_5_xhigh",
-    "mini_swe_agent_gpt_5_4_xhigh",
-    "mini_swe_agent_gpt_5_5_high",
-    "mini_swe_agent_gpt_5_5_low",
-    "mini_swe_agent_gpt_5_5_medium",
-    "mini_swe_agent_gpt_5_5_xhigh",
-    "mini_swe_agent_gpt_5_6_luna_high",
-    "mini_swe_agent_gpt_5_6_luna_low",
-    "mini_swe_agent_gpt_5_6_luna_max",
-    "mini_swe_agent_gpt_5_6_luna_medium",
-    "mini_swe_agent_gpt_5_6_luna_xhigh",
-    "mini_swe_agent_gpt_5_6_sol_high",
-    "mini_swe_agent_gpt_5_6_sol_low",
-    "mini_swe_agent_gpt_5_6_sol_max",
-    "mini_swe_agent_gpt_5_6_sol_medium",
-    "mini_swe_agent_gpt_5_6_sol_xhigh",
-    "mini_swe_agent_gpt_5_6_terra_high",
-    "mini_swe_agent_gpt_5_6_terra_low",
-    "mini_swe_agent_gpt_5_6_terra_max",
-    "mini_swe_agent_gpt_5_6_terra_medium",
-    "mini_swe_agent_gpt_5_6_terra_xhigh"
-  ],
-  "arm_spec": {
-    "mini_swe_agent_claude_fable_5_high": {
-      "model": "claude-fable-5",
-      "effort": "high",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-fable-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_claude_fable_5_low": {
-      "model": "claude-fable-5",
-      "effort": "low",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-fable-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_claude_fable_5_max": {
-      "model": "claude-fable-5",
-      "effort": "max",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-fable-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_claude_fable_5_medium": {
-      "model": "claude-fable-5",
-      "effort": "medium",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-fable-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_claude_fable_5_xhigh": {
-      "model": "claude-fable-5",
-      "effort": "xhigh",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-fable-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_4_8_high": {
-      "model": "claude-opus-4-8",
-      "effort": "high",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-4-8",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_4_8_low": {
-      "model": "claude-opus-4-8",
-      "effort": "low",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-4-8",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_4_8_max": {
-      "model": "claude-opus-4-8",
-      "effort": "max",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-4-8",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_4_8_medium": {
-      "model": "claude-opus-4-8",
-      "effort": "medium",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-4-8",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_4_8_xhigh": {
-      "model": "claude-opus-4-8",
-      "effort": "xhigh",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-4-8",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_5_high": {
-      "model": "claude-opus-5",
-      "effort": "high",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_5_low": {
-      "model": "claude-opus-5",
-      "effort": "low",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_5_max": {
-      "model": "claude-opus-5",
-      "effort": "max",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_5_medium": {
-      "model": "claude-opus-5",
-      "effort": "medium",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_claude_opus_5_xhigh": {
-      "model": "claude-opus-5",
-      "effort": "xhigh",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-opus-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_4_6_high": {
-      "model": "claude-sonnet-4-6",
-      "effort": "high",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-4-6",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_5_high": {
-      "model": "claude-sonnet-5",
-      "effort": "high",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_5_low": {
-      "model": "claude-sonnet-5",
-      "effort": "low",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_5_max": {
-      "model": "claude-sonnet-5",
-      "effort": "max",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_5_medium": {
-      "model": "claude-sonnet-5",
-      "effort": "medium",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_claude_sonnet_5_xhigh": {
-      "model": "claude-sonnet-5",
-      "effort": "xhigh",
-      "provider": "anthropic",
-      "request_kwargs": {
-        "model": "claude-sonnet-5",
-        "thinking": {
-          "type": "adaptive"
-        },
-        "output_config": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_4_xhigh": {
-      "model": "gpt-5.4",
-      "effort": "xhigh",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.4",
-        "reasoning": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_5_high": {
-      "model": "gpt-5.5",
-      "effort": "high",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.5",
-        "reasoning": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_5_low": {
-      "model": "gpt-5.5",
-      "effort": "low",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.5",
-        "reasoning": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_5_medium": {
-      "model": "gpt-5.5",
-      "effort": "medium",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.5",
-        "reasoning": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_5_xhigh": {
-      "model": "gpt-5.5",
-      "effort": "xhigh",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.5",
-        "reasoning": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_luna_high": {
-      "model": "gpt-5.6-luna",
-      "effort": "high",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-luna",
-        "reasoning": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_luna_low": {
-      "model": "gpt-5.6-luna",
-      "effort": "low",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-luna",
-        "reasoning": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_luna_max": {
-      "model": "gpt-5.6-luna",
-      "effort": "max",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-luna",
-        "reasoning": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_luna_medium": {
-      "model": "gpt-5.6-luna",
-      "effort": "medium",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-luna",
-        "reasoning": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_luna_xhigh": {
-      "model": "gpt-5.6-luna",
-      "effort": "xhigh",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-luna",
-        "reasoning": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_sol_high": {
-      "model": "gpt-5.6-sol",
-      "effort": "high",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-sol",
-        "reasoning": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_sol_low": {
-      "model": "gpt-5.6-sol",
-      "effort": "low",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-sol",
-        "reasoning": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_sol_max": {
-      "model": "gpt-5.6-sol",
-      "effort": "max",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-sol",
-        "reasoning": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_sol_medium": {
-      "model": "gpt-5.6-sol",
-      "effort": "medium",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-sol",
-        "reasoning": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_sol_xhigh": {
-      "model": "gpt-5.6-sol",
-      "effort": "xhigh",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-sol",
-        "reasoning": {
-          "effort": "xhigh"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_terra_high": {
-      "model": "gpt-5.6-terra",
-      "effort": "high",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-terra",
-        "reasoning": {
-          "effort": "high"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_terra_low": {
-      "model": "gpt-5.6-terra",
-      "effort": "low",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-terra",
-        "reasoning": {
-          "effort": "low"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_terra_max": {
-      "model": "gpt-5.6-terra",
-      "effort": "max",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-terra",
-        "reasoning": {
-          "effort": "max"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_terra_medium": {
-      "model": "gpt-5.6-terra",
-      "effort": "medium",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-terra",
-        "reasoning": {
-          "effort": "medium"
-        }
-      }
-    },
-    "mini_swe_agent_gpt_5_6_terra_xhigh": {
-      "model": "gpt-5.6-terra",
-      "effort": "xhigh",
-      "provider": "openai",
-      "request_kwargs": {
-        "model": "gpt-5.6-terra",
-        "reasoning": {
-          "effort": "xhigh"
-        }
-      }
-    }
-  },
-  "k": 12,
-  "tau": 0.5,
-  "sim_floor": 0.35,
-  "fallback_arm_index": 10,
-  "provenance": "Same DeepSWE v1.1 supervision as the cloud-embedded router_v0.json (41 arms x 110 tasks over 88 repos), but embedded locally with mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ instead of OpenAI. A separate 80/20 repo-split holdout (not the 5-fold CV below) found LOCAL embeddings gave cost ratio median 3.79x vs OpenAI's 3.18x, graded-delta median -0.021 vs -0.015, across 6 seeds -- comparable, not yet validated at the same rigor as the cloud variant's nested-CV headline number. | embeddings computed in-process (mlx_embeddings on Apple Silicon; sentence-transformers Qwen/Qwen3-Embedding-0.6B elsewhere)",
-  "scope_warning": "INPUT SHAPE MATTERS. Fit on repo-issue statements of p10=955 / p50=1976 / p90=3450 characters. In-distribution nearest-neighbour cosine similarity runs min 0.266 / p50 0.439; sim_floor=0.35 is the p10, so it abstains on the least-covered ~10% of in-distribution tasks. Short one-line prompts score 0.14-0.34 and will ALWAYS abstain to the strongest arm -- safe, but no saving. Long-horizon tasks only (median 61 agent steps). NOT validated on short interactive requests or on any held-out benchmark. Always check Decision.off_distribution before trusting a route.",
-  "measured_artifact_behaviour": "Self-test numbers below (this artifact, re-run under the same repo-grouped folds) are the only measured claim for this variant -- no separate ungated-vs-guarded comparison has been run yet, unlike the cloud variant.",
-  "n_tasks": 110,
-  "n_repos": 88,
-  "embed_backend": "local"
 }

 {
+ "version": "v1-trained",
+ "kind": "trained",
+ "hf_repo": "experiential-labs/coding-router",
+ "embed_model_mlx": "encoder-mlx-4bit",
+ "embed_model_torch": "encoder-fp16",
+ "T": 0.05453708011655199,
+ "lam": 0.01,
+ "sim_floor": 0.459,
+ "fallback_arm_index": 10,
+ "arms": [
+  "mini_swe_agent_claude_fable_5_high",
+  "mini_swe_agent_claude_fable_5_low",
+  "mini_swe_agent_claude_fable_5_max",
+  "mini_swe_agent_claude_fable_5_medium",
+  "mini_swe_agent_claude_fable_5_xhigh",
+  "mini_swe_agent_claude_opus_4_8_high",
+  "mini_swe_agent_claude_opus_4_8_low",
+  "mini_swe_agent_claude_opus_4_8_max",
+  "mini_swe_agent_claude_opus_4_8_medium",
+  "mini_swe_agent_claude_opus_4_8_xhigh",
+  "mini_swe_agent_claude_opus_5_high",
+  "mini_swe_agent_claude_opus_5_low",
+  "mini_swe_agent_claude_opus_5_max",
+  "mini_swe_agent_claude_opus_5_medium",
+  "mini_swe_agent_claude_opus_5_xhigh",
+  "mini_swe_agent_claude_sonnet_4_6_high",
+  "mini_swe_agent_claude_sonnet_5_high",
+  "mini_swe_agent_claude_sonnet_5_low",
+  "mini_swe_agent_claude_sonnet_5_max",
+  "mini_swe_agent_claude_sonnet_5_medium",
+  "mini_swe_agent_claude_sonnet_5_xhigh",
+  "mini_swe_agent_gpt_5_4_xhigh",
+  "mini_swe_agent_gpt_5_5_high",
+  "mini_swe_agent_gpt_5_5_low",
+  "mini_swe_agent_gpt_5_5_medium",
+  "mini_swe_agent_gpt_5_5_xhigh",
+  "mini_swe_agent_gpt_5_6_luna_high",
+  "mini_swe_agent_gpt_5_6_luna_low",
+  "mini_swe_agent_gpt_5_6_luna_max",
+  "mini_swe_agent_gpt_5_6_luna_medium",
+  "mini_swe_agent_gpt_5_6_luna_xhigh",
+  "mini_swe_agent_gpt_5_6_sol_high",
+  "mini_swe_agent_gpt_5_6_sol_low",
+  "mini_swe_agent_gpt_5_6_sol_max",
+  "mini_swe_agent_gpt_5_6_sol_medium",
+  "mini_swe_agent_gpt_5_6_sol_xhigh",
+  "mini_swe_agent_gpt_5_6_terra_high",
+  "mini_swe_agent_gpt_5_6_terra_low",
+  "mini_swe_agent_gpt_5_6_terra_max",
+  "mini_swe_agent_gpt_5_6_terra_medium",
+  "mini_swe_agent_gpt_5_6_terra_xhigh"
+ ],
+ "arm_spec": {
+  "mini_swe_agent_claude_fable_5_high": {
+   "model": "claude-fable-5",
+   "effort": "high",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-fable-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_claude_fable_5_low": {
+   "model": "claude-fable-5",
+   "effort": "low",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-fable-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_claude_fable_5_max": {
+   "model": "claude-fable-5",
+   "effort": "max",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-fable-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_claude_fable_5_medium": {
+   "model": "claude-fable-5",
+   "effort": "medium",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-fable-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_claude_fable_5_xhigh": {
+   "model": "claude-fable-5",
+   "effort": "xhigh",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-fable-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_4_8_high": {
+   "model": "claude-opus-4-8",
+   "effort": "high",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-4-8",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_4_8_low": {
+   "model": "claude-opus-4-8",
+   "effort": "low",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-4-8",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_4_8_max": {
+   "model": "claude-opus-4-8",
+   "effort": "max",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-4-8",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_4_8_medium": {
+   "model": "claude-opus-4-8",
+   "effort": "medium",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-4-8",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_4_8_xhigh": {
+   "model": "claude-opus-4-8",
+   "effort": "xhigh",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-4-8",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_5_high": {
+   "model": "claude-opus-5",
+   "effort": "high",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_5_low": {
+   "model": "claude-opus-5",
+   "effort": "low",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_5_max": {
+   "model": "claude-opus-5",
+   "effort": "max",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_5_medium": {
+   "model": "claude-opus-5",
+   "effort": "medium",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_claude_opus_5_xhigh": {
+   "model": "claude-opus-5",
+   "effort": "xhigh",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-opus-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_4_6_high": {
+   "model": "claude-sonnet-4-6",
+   "effort": "high",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-4-6",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_5_high": {
+   "model": "claude-sonnet-5",
+   "effort": "high",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_5_low": {
+   "model": "claude-sonnet-5",
+   "effort": "low",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_5_max": {
+   "model": "claude-sonnet-5",
+   "effort": "max",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_5_medium": {
+   "model": "claude-sonnet-5",
+   "effort": "medium",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_claude_sonnet_5_xhigh": {
+   "model": "claude-sonnet-5",
+   "effort": "xhigh",
+   "provider": "anthropic",
+   "request_kwargs": {
+    "model": "claude-sonnet-5",
+    "thinking": {
+     "type": "adaptive"
+    },
+    "output_config": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_4_xhigh": {
+   "model": "gpt-5.4",
+   "effort": "xhigh",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.4",
+    "reasoning": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_5_high": {
+   "model": "gpt-5.5",
+   "effort": "high",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.5",
+    "reasoning": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_5_low": {
+   "model": "gpt-5.5",
+   "effort": "low",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.5",
+    "reasoning": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_5_medium": {
+   "model": "gpt-5.5",
+   "effort": "medium",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.5",
+    "reasoning": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_5_xhigh": {
+   "model": "gpt-5.5",
+   "effort": "xhigh",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.5",
+    "reasoning": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_luna_high": {
+   "model": "gpt-5.6-luna",
+   "effort": "high",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-luna",
+    "reasoning": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_luna_low": {
+   "model": "gpt-5.6-luna",
+   "effort": "low",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-luna",
+    "reasoning": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_luna_max": {
+   "model": "gpt-5.6-luna",
+   "effort": "max",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-luna",
+    "reasoning": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_luna_medium": {
+   "model": "gpt-5.6-luna",
+   "effort": "medium",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-luna",
+    "reasoning": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_luna_xhigh": {
+   "model": "gpt-5.6-luna",
+   "effort": "xhigh",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-luna",
+    "reasoning": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_sol_high": {
+   "model": "gpt-5.6-sol",
+   "effort": "high",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-sol",
+    "reasoning": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_sol_low": {
+   "model": "gpt-5.6-sol",
+   "effort": "low",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-sol",
+    "reasoning": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_sol_max": {
+   "model": "gpt-5.6-sol",
+   "effort": "max",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-sol",
+    "reasoning": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_sol_medium": {
+   "model": "gpt-5.6-sol",
+   "effort": "medium",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-sol",
+    "reasoning": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_sol_xhigh": {
+   "model": "gpt-5.6-sol",
+   "effort": "xhigh",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-sol",
+    "reasoning": {
+     "effort": "xhigh"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_terra_high": {
+   "model": "gpt-5.6-terra",
+   "effort": "high",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-terra",
+    "reasoning": {
+     "effort": "high"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_terra_low": {
+   "model": "gpt-5.6-terra",
+   "effort": "low",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-terra",
+    "reasoning": {
+     "effort": "low"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_terra_max": {
+   "model": "gpt-5.6-terra",
+   "effort": "max",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-terra",
+    "reasoning": {
+     "effort": "max"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_terra_medium": {
+   "model": "gpt-5.6-terra",
+   "effort": "medium",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-terra",
+    "reasoning": {
+     "effort": "medium"
+    }
+   }
+  },
+  "mini_swe_agent_gpt_5_6_terra_xhigh": {
+   "model": "gpt-5.6-terra",
+   "effort": "xhigh",
+   "provider": "openai",
+   "request_kwargs": {
+    "model": "gpt-5.6-terra",
+    "reasoning": {
+     "effort": "xhigh"
+    }
+   }
+  }
+ },
+ "n_tasks": 110,
+ "n_repos": 88,
+ "provenance": "EXP-012 winning recipe `reward_lcb_b0.2`: Qwen3-Embedding-0.6B with LoRA r=16 (q/k/v/o projections) trained by exact expected reward -E_pi[graded - 3.0*cost] on the LiveCodeBench 7x76 matrix, KL(pi||pi_init) anchor beta=0.2, 150 steps, checkpoint step 50 (T=0.0545), lam=0.01 -- (step, lam) selected ONLY on the six standard seeds' inner 75/25 DeepSWE-train splits (feasibility inner_graded >= inner-best-arm - 0.01, then max mean cost ratio; feasible on 5/6 inner splits). Memory/calibration side: the full 110-task DeepSWE v1.1 evidence (41 arms x 110 tasks, 88 repos), bank embedded with the merged fp16 encoder via sentence-transformers. Sweep evidence (per-seed selection, 6 seeds): selected-holdout mean graded 0.9484 at $60.17/split -- +0.015 graded and 2.1x cheaper than the deployable always-best-train baseline (0.9336/$126.28); parity, NOT better, vs the hindsight-best static arm. Cost figures are matrix-based (June 2026 collection) and pending live re-benchmark per EXP-014 (live drift measured).",
+ "scope_warning": "INPUT SHAPE MATTERS. The decision rule was validated on repo-issue statements (p50 ~2,000 chars, long-horizon SWE tasks, median 61 agent steps). The encoder was tuned on LCB competitive-programming tasks and evaluated ONLY on the DeepSWE holdout; sim_floor abstains (escalates to the strongest arm) when no bank task is close in the TUNED space. Short one-line prompts will usually abstain -- safe, but no saving. Always check Decision.off_distribution.",
+ "embed_backend": "local",
+ "selection": {
+  "feas_count": 5,
+  "mean_inner_ratio": 2.4618640944304526,
+  "mean_inner_graded": 0.9208466112613678
+ }
 }

router.npz CHANGED Viewed

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9b6a0299f4bbd5adc2136df826e629f4336cb8f9503a2cdeafe623e9a80bd612
-size 275358

 version https://git-lfs.github.com/spec/v1
+oid sha256:29b71850c753cfc04f5dc992cdaf8e9eb4211388b8cb652fd0df21243ac91303
+size 274172