Upload qwen4b taskvector merge

.gitattributes

@@ -33,3 +33,4 @@ 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
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 2024 Alibaba Cloud
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

MANIFEST.sha256

@@ -0,0 +1,20 @@
+832dd9e00a68dd83b3c3fb9f5588dad7dcf337a0db50f7d9483f310cd292e92e  LICENSE
+5e66b33663c05d41fe3758efd5dde9aaa83e587e032e6f4d458a6fc74bc881f9  OFFICIAL_MERGE_RECEIPT.json
+1592a771cd3fe8b9e6a976c36a2f8edbcd8c373b4ff5645c69911f0944802f33  README.md
+490367bf1e2bd55f5a7381dc4083bcffc18430bcbc26457680c5c17268225bcc  adapter/MANIFEST.json
+16a5f02dafeca4c7804f7306e353375ac567d8158ed37baa20b80dcfe9188f33  adapter/TOKEN_ROWS_META.json
+4e04ab4bb3a660f837b6f57d3964cebfdf5b5a58cebc0c5cd9a30372a396212c  adapter/adapter_config.json
+619dcee8c5c04416fed0cee928375583332ce2986bf3f51c0caf8408433831ed  adapter/adapter_model.safetensors
+407274bf3280d79ed8efabaf3c3b3e75debaeb6b8efdb4433ac0bae934144688  adapter/token_rows_both_sides.safetensors
+87a2728cb8dc9fe424d624542f6060ec05a1d285ebbec578bb078900e33396b5  chat_template.jinja
+40bcf4e0a6a7b5eb8aa582ee23f0c9700d76f147d17159419893bbad0f762176  config.json
+a2ac0d07aac37502207a212b122abe32bd9564266594a317c0ff490ebce1f226  generation_config.json
+0e43b8415e406761c7ab73c8225fde51abdb7f51a4b80a9e90c616f53c3ea419  model.safetensors
+5ebf618c3d547252a8078665415fa4fb3b080d071c7e0f2750f6af44e231bc3e  provenance/DEV256_COMPLETE.json
+24fc75c1ce08f0e4b7dcc31684c36c1f95b7032999c0e40308426781728f939c  provenance/MQ0_NEAR_DUP.json
+c9f6e3b33a4693354da05fd6f18b3d05932430ca7c8dc9f7f008d3dcc49ac592  provenance/POLICY.json
+5c45fca93b73d0293c39913fc74160cc24f3ed5efa4bfd51582428cc029e1cd2  provenance/RUN_IDENTITY.json
+0304abe092d416ca7974b8f98f39783c52e71467bea8382ed000315ed917ec8a  provenance/TRAINING_CONFIG.json
+2ff0f61e200d18dd55b1301f9573e923abd6fe87608220f7126cf709d4cb7288  provenance/build_mix_distill_payload.py
+be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506  tokenizer.json
+1749ac66e3e7ce3862337c23cca5895a2ee131929563e18a944f8c1c80363712  tokenizer_config.json

OFFICIAL_MERGE_RECEIPT.json

@@ -0,0 +1 @@
+{"adapter_tree_sha256":"467d7c210b6001b8a1c638cce88ebb5112fa2f7d1684000f712b4ccafa814b29","base_model_path":"/workspace/code-sft-infra/models/qwen3-4b-base","checkpoint_manifest_sha256":"490367bf1e2bd55f5a7381dc4083bcffc18430bcbc26457680c5c17268225bcc","checkpoint_path":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450","label":"v4-q4b-think-mix","merge_backend":"merge_for_eval","merged_at_unix":1788962266.3138773,"merged_model_path":"/workspace/code-sft-runs/rstar-eval/official-dev256/v4-q4b-think-mix/scratch/step-000920-tokens-64848450-merged-full","merged_tree_sha256":"4d14b1af017f7e5f113017a3a2e1a29942463afde5eca0d4f4f9faa0c17194ae","receipt_sha256":"defb397b004f96a114242399a45d919ecdc2c88fc3fde9e409e425bac453af12","schema":"CODE_SFT_OFFICIAL_DEV256_MERGE_RECEIPT_V1","token_rows_meta_path":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450/TOKEN_ROWS_META.json","token_rows_meta_worker":"rstar_dual_worker_v4","untied_export":{"lm_head_weight_shard":"model.safetensors","tie_word_embeddings":false}}

README.md

@@ -0,0 +1,29 @@
+---
+license: apache-2.0
+library_name: transformers
+base_model: Qwen/Qwen3-4B-Base
+tags:
+- qwen3
+- merge
+- taskvector
+- code
+- think
+---
+
+# Montalte/qwen4b-code-think-taskvector
+
+Unified Qwen3-4B merge artifact for directional math↔code transfer experiments.
+
+| Field | Value |
+|-------|-------|
+| Base | `Qwen/Qwen3-4B-Base@906bfd4b4dc7f14ee4320094d8b41684abff8539` |
+| Source specialist | `modrill/code-think-q4b-20260908` |
+| Domain | code |
+| Mode | think |
+| Method | **taskvector** |
+
+## Method
+
+**taskvector** = dense source specialist \(\theta_s\) (full SFT weights; equivalent to applying the complete task vector \(\tau_s=\theta_s-\theta_0\)).
+
+Uploaded: 2026-09-11.

chat_template.jinja

@@ -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 %}

config.json

@@ -0,0 +1,71 @@
+{
+  "architectures": [
+    "Qwen3ForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 151643,
+  "dtype": "bfloat16",
+  "eos_token_id": 151643,
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 2560,
+  "initializer_range": 0.02,
+  "intermediate_size": 9728,
+  "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",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention",
+    "full_attention"
+  ],
+  "max_position_embeddings": 32768,
+  "max_window_layers": 36,
+  "model_type": "qwen3",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 36,
+  "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": false,
+  "transformers_version": "5.16.1",
+  "use_cache": true,
+  "use_sliding_window": false,
+  "vocab_size": 151936
+}

generation_config.json

@@ -0,0 +1,7 @@
+{
+  "bos_token_id": 151643,
+  "do_sample": false,
+  "eos_token_id": 151643,
+  "max_new_tokens": 2048,
+  "transformers_version": "5.16.1"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e43b8415e406761c7ab73c8225fde51abdb7f51a4b80a9e90c616f53c3ea419
+size 8822894520

provenance/DEV256_COMPLETE.json

@@ -0,0 +1 @@
+{"adapter":{"adapter_identity":"MERGED_FULL_WEIGHTS_NO_LORA","adapter_path":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450/adapter","adapter_tree_sha256":"467d7c210b6001b8a1c638cce88ebb5112fa2f7d1684000f712b4ccafa814b29","checkpoint":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450","checkpoint_manifest_sha256":"490367bf1e2bd55f5a7381dc4083bcffc18430bcbc26457680c5c17268225bcc","merged_path":"/workspace/code-sft-runs/rstar-eval/official-dev256/v4-q4b-think-mix/scratch/step-000920-tokens-64848450-merged-full"},"arm":"official-v4-q4b-think-mix","base":false,"benchmark_contract":{"date_filter":null,"native_index":{"bytes":2546647101,"immutable_object_uri":"r2mtx:mxx/isolated-backups/full-workspace/20260826T031752Z-8bcbe83efc6607702f6b681b4429d72dc6c1a4c570f18c85e8c7cee5743c682a/objects/sha256/ea/ea66e0d8814490a7e0d7ff555297e43a7d9662405c16609f9df51b42ec1b1cc1","path":"/workspace/code-sft-runs/rstar-eval/contracts/dev256/native_index.jsonl","sha256":"ea66e0d8814490a7e0d7ff555297e43a7d9662405c16609f9df51b42ec1b1cc1"},"ordered_task_ids_sha256":"8f4b14d4af19399ed59a052ef970ad18eb83f19cc5257bcafa4fafd1938315b9","prompts":{"path":"/workspace/code-sft-runs/rstar-eval/contracts/dev256/evalscope_prompts.jsonl","sha256":"b8c2047bf804c01a1cc82d16e07599dea0f4d94a42af6a26386c832a33747695"},"recovery":{"path":"/workspace/code-sft-runs/rstar-eval/contracts/dev256/RECOVERY.json","sha256":"011d8adba4986651a5d83add11626f155270008c9e69e89a4a86e753d71b0906"},"role":"DIAGNOSTIC_ONLY","rows":256,"schema":"CTX_LCDIV_DEV256_RECOVERY_V2","seed":3407,"seed_count":1,"subset_filter":null,"universe_rule":"exact content-addressed ordered task IDs recovered from six frozen test*.jsonl shards; no resampling, shuffle, date filtering, subset filtering, or hub refresh"},"cap_rate":0.53125,"caps":136,"checkpoint":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450","code_plus_stop":73,"first_stop_accounted":256,"first_stop_token":{"100257":0,"100265":0,"151643":0,"151645":0,"length":136,"other":120},"gpu":2,"kind":"merged_full_weights","lora":false,"merged":"/workspace/code-sft-runs/rstar-eval/official-dev256/v4-q4b-think-mix/scratch/step-000920-tokens-64848450-merged-full","mode":"think","native_validation":{"modes":{"functional":109,"stdin":147},"native_index_sha256":"ea66e0d8814490a7e0d7ff555297e43a7d9662405c16609f9df51b42ec1b1cc1","ordered_evaluation_samples_sha256":"811cc4cd88b1863ab973be0b906c12934f1acc67d436f1b48d4a6863f7e0a4c7","ordered_task_ids_sha256":"8f4b14d4af19399ed59a052ef970ad18eb83f19cc5257bcafa4fafd1938315b9","rows":256},"pass_at_1":0.28515625,"passed":73,"pending_infra":0,"per_item":{"document_sha256":"ea2d8faff8649d26c33a83b40fea63e74221e4ce76f22c052259e721495b4b8b","path":"/workspace/code-sft-runs/rstar-eval/official-dev256/v4-q4b-think-mix/results/PER_ITEM.json","sha256":"c21c2587b33b0c6d2ed14da02c463fc26e57a3c64456f4e6c82c9ca8a45cc98a"},"port":18617,"protocol_failure":0,"queue_sha256":"80a75fca2a228428ecac539eb773958b8c38975d9adb727093548aefa27fa39b","receipt_sha256":"2e587612516917f00ff268d1309f1533cb6cb282dc1b25b4ed3170517df70036","renderer_authority":{"path":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/tokenizers/qwen3-v4-tail151643.authority.json","sha256":"6b1461404d033b79f05a8cbb373f05a495d506a567b35b18cbf89930f345422a"},"renderer_id":"qwen3-8b-think-lcb-system-v2","rows":256,"runner":{"evalscope_adapter_sha256":"6c4977439b751003d6900c5cec2e7b3244599d3427903f99e3d7300d2c0d7687","evalscope_core_files":{"evaluate_utils.py":"a16167430f636906d15c6540f4640bb72973d9c34cea87928ebe870b51fc9794","extract_utils.py":"1c74c40a9250191cdaa5560104df583a4bec493154a2ea3eb603c35237c2a038","load_utils.py":"7c4cd6cff7d41b08b935cba6e4f35517d8c572e0f01a0b3b56dff5d182811364","testing_util.py":"72ac3f87e288cf39ed882e6f31cfc1192e5027d97d7d9a7730e2aa946effa329"},"evalscope_sandbox_sha256":"fcbbd6ec7634225ce023da194d178206ef778ad8e945e0ac36353934c09186f0","evalscope_version":"1.11.0","library_sha256":"0ab605fc8f758b1507752f19b4266c0a17e562774b1dcbe8e138708fd76ad7df","official_task":{"base":false,"cache_root":"/workspace/code-sft-runs/rstar-eval/official-dev256/v4-q4b-think-mix/cache","checkpoint":"/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out/Q4B-THINK-MIX/906bfd4b4dc7f14ee4320094d8b41684abff8539/qwen3-8b-think-lcb-system-v2/fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885/step-000920-tokens-64848450","label":"v4-q4b-think-mix","mode":"think","student_family":"qwen"},"runner_id":"code-sft-rstar-lcb-handoff-v2","runner_sha256":"5ea2f032337a72c97f0c8cda2310e9f99e649f2bbe8cbcea8e2bcbd5072c6431","sandbox_runtime":{"execution_module_sha256":"cb162f70bb548694ad08372c71399b9011500b198cc7d23bf5dc2f95912189fc","pool_config_sha256":"33de3626e6591d5dcb0b1b99bb8fa53b05acd42a8a9fe65637e2fd2c890b0790","pool_probe_sha256":"6b2a91c4f5ff066ca9617602d166f4e19a9d5228a990ea2aa62b6fd76959ebc0","scope_lease":"CODE_SFT_SANDBOX_SCOPE_LEASE_V1","scope_module_sha256":"d5c0cb716af86aa2bc03c12cce8106046692672eafe6fb7cc7562e4c9752a8bf","transport":"SANDBOX_API_V1_DIRECT_TOOL_EXECUTE"}},"schema":"CODE_SFT_RSTAR_EVAL_SUITE_RECEIPT_V1","scored":256,"seed":3407,"served_model":"v4-q4b-think-mix-dev256","status":"COMPLETE","student_family":"qwen","suite":"dev256","unscored":0}

provenance/MQ0_NEAR_DUP.json

@@ -0,0 +1,13 @@
+{
+  "mq0": {
+    "status": "PASS",
+    "coverage": "6_OF_6",
+    "intersection": 0,
+    "denylist_path": "/workspace/code_selfgen_rstar_20260831_t4b2507/ext/work/BENCHMARK_HASHES.txt",
+    "denylist_sha256": "c6ed508202aa5a3d1eb3a49f1791e241e02a8f99e7a81bc4ba1a4164d6f1c2e9"
+  },
+  "near_dup_hit_hashes": [],
+  "near_dup_grey_zone_count": 0,
+  "dropped_n": 0,
+  "kept_n": 4155
+}

provenance/POLICY.json

@@ -0,0 +1,169 @@
+{
+  "arms": [
+    {
+      "arm_id": "Q4B-THINK-MIX",
+      "data": {
+        "assistant_target_tokens": 64848450,
+        "manifest_file_sha256": "fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885",
+        "manifest_path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/mix_l0.2_s0.8_tail151643_2EP_READY.json",
+        "manifest_sha256": "fc22e609554e898f2ae570f75bce208ae3428aada9931dc43bc7da3735f95885",
+        "max_assistant_target_tokens_per_example": 16384,
+        "payload_rows": 8310,
+        "payload_sha256": "6097d7bf8bc18621ed4fbc2c7f4e6a68b65186eb1d25964fd632580f97dabdfd",
+        "source_revision": "3a7a0a0636ec96e3c1ec42ebe79ade467caa040d"
+      },
+      "gpu": 2,
+      "model": {
+        "config_sha256": "304b2545a258d35620f1d4bf46940c0471d9baa00715ff8e77f84c2fca5057c1",
+        "hf_id": "Qwen/Qwen3-4B-Base",
+        "local_path": "/workspace/code-sft-infra/models/qwen3-4b-base",
+        "revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+        "tie_word_embeddings": true,
+        "tokenizer_sha256": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539"
+      },
+      "model_id": "qwen3-4b-base",
+      "renderer": {
+        "assistant_termination": {
+          "final_text_suffix": "<|endoftext|>",
+          "final_token_suffix": [
+            151643
+          ],
+          "turn_end_token_id": 151645
+        },
+        "chat_template_kwargs": {
+          "enable_thinking": true
+        },
+        "chat_template_sha256": "b608d01889daca5f325cc44d1cbabf73ec74d4d53c9950725cc46068a7006a68",
+        "encoding_policy": {
+          "add_special_tokens": false,
+          "assistant_termination_in_loss": true,
+          "assistant_termination_in_target_census": true,
+          "incomplete_target": "reject",
+          "padding_label_mask": "attention_or_pad_positions_never_token_id"
+        },
+        "eos_token": "<|endoftext|>",
+        "eos_token_id": 151643,
+        "mode": "think",
+        "runtime_tokenizer_manifest_sha256": "66615d472270d50fca0a4b4b299ab2663a9944102f0afd4626fcee46214fe248",
+        "runtime_tokenizer_path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/tokenizers/qwen3-v4-tail151643",
+        "stop_token_ids": [
+          151643,
+          151645
+        ],
+        "system_prompt": "You are an expert Python programmer. You will be given a question (problem specification) and will generate a correct Python program that matches the specification and passes all tests. You will NOT return anything except for the program.",
+        "system_prompt_policy": "exact_evalscope_live_code_bench_system_message",
+        "system_prompt_policy_sha256": "209896b91040ef82e511aa0c7c95a2f9f5dadc9d3c4879f7c66da4ddca34295b",
+        "system_prompt_sha256": "41b9af0ebedb936e74e0996175f76f0116c334375ec3a7c914166f126c02f489",
+        "tokenizer_hf_id": "Qwen/Qwen3-4B-Base",
+        "tokenizer_revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+        "tokenizer_sha256": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539"
+      },
+      "renderer_authority": {
+        "path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/tokenizers/qwen3-v4-tail151643.authority.json",
+        "sha256": "6b1461404d033b79f05a8cbb373f05a495d506a567b35b18cbf89930f345422a"
+      },
+      "renderer_id": "qwen3-8b-think-lcb-system-v2",
+      "source_id": "rstar-coder"
+    }
+  ],
+  "generation_behavior_gate": null,
+  "owner_authorization": "PI 2026-09-09 Q4B-THINK-MIX Mix Distillation 0.2:0.8",
+  "recipe": {
+    "adapter": {
+      "alpha": 128,
+      "dropout": 0.0,
+      "method": "lora",
+      "modules_to_save": null,
+      "rank": 64,
+      "target_modules": [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj"
+      ]
+    },
+    "boundary": {
+      "fp32_master": true,
+      "mode": "trainable_token_rows_both_sides",
+      "row_lr": 0.0001,
+      "row_weight_decay": 0.0,
+      "token_ids": [
+        151643,
+        151667,
+        151668
+      ],
+      "untie_tied_embeddings": true
+    },
+    "checkpoint_target_tokens": [
+      32445505,
+      48759220,
+      64848450
+    ],
+    "epochs": 2,
+    "gradient_checkpointing": true,
+    "max_sequence_length": 32768,
+    "model_runtime": {
+      "bf16": true,
+      "finetuning_type": "lora",
+      "gradient_checkpointing": true,
+      "lora_alpha": 128,
+      "lora_dropout": 0.0,
+      "lora_rank": 64,
+      "lora_targets": [
+        "q_proj",
+        "k_proj",
+        "v_proj",
+        "o_proj",
+        "gate_proj",
+        "up_proj",
+        "down_proj"
+      ],
+      "modules_to_save": null,
+      "trainable_dtype": "float32",
+      "trainable_token_indices": {
+        "embed_tokens": [
+          151643,
+          151667,
+          151668
+        ],
+        "lm_head": [
+          151643,
+          151667,
+          151668
+        ]
+      }
+    },
+    "optimizer": {
+      "betas": [
+        0.9,
+        0.95
+      ],
+      "eps": 1e-08,
+      "gradient_clip_norm": 1.0,
+      "learning_rate": 0.0001,
+      "name": "adamw",
+      "weight_decay": 0.1
+    },
+    "packing": {
+      "enabled": false
+    },
+    "precision": "bfloat16",
+    "repeat_examples": false,
+    "scheduler": "cosine_by_assistant_target_token_dose",
+    "seed": 42,
+    "target_assistant_tokens": 64848450,
+    "tokens_per_optimizer_update": 65536,
+    "trainable_dtype": "float32",
+    "truncation": false,
+    "warmup_target_tokens": 3890907
+  },
+  "runtime": {
+    "checkpoint_root": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/out",
+    "state_root": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX"
+  },
+  "schema": "CODE_SFT_RSTAR_DUAL_TRAINING_POLICY_V1",
+  "status": "SEALED"
+}

provenance/RUN_IDENTITY.json

@@ -0,0 +1,123 @@
+{
+  "schema": "RUN_IDENTITY_V1",
+  "run_id": "t30b2507-q4b-think-mix-l0.2-s0.8-v4-tail151643",
+  "corrects": "Q4B-THINK Mix Distillation 0.2:0.8 (30B-A3B-Thinking : 4B-Thinking by rows)",
+  "notes": "Mix Distillation Mix-Large (Li et al. 2025 arXiv 2502.12143): 30B-A3B-Thinking : 4B-Thinking = 0.2 : 0.8 by row/problem; one teacher trace per problem; dose target 32436894/ep ±1%, actual 1ep 32424225, 2ep 64848450. 30B source=/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/data/THINK_PAIRED_V4_2EP.parquet sha=c160713fea76eae124dbbc3b5c619d3164862da83f61d91d3970ede990a12540 (2EP first half 4715 rows). 4B source=/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-T4BDATA/data/joint-full-t4b_tail151643.parquet sha=d09c4d6272de7e50c9b2df0e401878864f09bda33db05a5833c43575969fa524. V4 recipe unchanged (LoRA r64/α128, lr 1e-4, both-sides B-row [151643,151667,151668], physical 2ep concat, seed 42). PI approved 2026-09-09.",
+  "model": {
+    "hf_id": "Qwen/Qwen3-4B-Base",
+    "revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+    "local_path": "/workspace/code-sft-infra/models/qwen3-4b-base",
+    "tied_embeddings": true
+  },
+  "data": {
+    "path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/mix_l0.2_s0.8_tail151643_2EP.parquet",
+    "rows": 8310,
+    "sha256": "6097d7bf8bc18621ed4fbc2c7f4e6a68b65186eb1d25964fd632580f97dabdfd",
+    "census": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/TOKEN_CENSUS_MIX.json",
+    "physical_2ep": "concat(mix, mix) same order; teacher_source 30b/4b; tail 151643",
+    "source_1ep_path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/mix_l0.2_s0.8_tail151643.parquet",
+    "source_1ep_sha256": "3c5ae286e7e5242df3c011123edaf0727405d3b535e85d77ce6b99afe7c27e77",
+    "source_1ep_rows": 4155,
+    "teacher_30b_path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/data/THINK_PAIRED_V4_2EP.parquet",
+    "teacher_30b_sha256": "c160713fea76eae124dbbc3b5c619d3164862da83f61d91d3970ede990a12540",
+    "teacher_4b_path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-T4BDATA/data/joint-full-t4b_tail151643.parquet",
+    "teacher_4b_sha256": "d09c4d6272de7e50c9b2df0e401878864f09bda33db05a5833c43575969fa524",
+    "mix_ratio_rows": "0.2:0.8",
+    "n30": 831,
+    "n4": 3324,
+    "active_tokens_1ep": 32424225
+  },
+  "seeds": {
+    "train": 42,
+    "eval": 3407,
+    "mix": 42
+  },
+  "parameters": {
+    "epochs": 2,
+    "target_assistant_tokens": 64848450,
+    "active_tokens_1ep": 32424225,
+    "checkpoint_target_tokens": [
+      32445505,
+      48759220,
+      64848450
+    ],
+    "milestone_updates": [
+      495,
+      744,
+      990
+    ],
+    "warmup_target_tokens": 3890907,
+    "learning_rate": 0.0001,
+    "row_lr": 0.0001,
+    "optimizer": "adamw",
+    "optimizer_betas": [
+      0.9,
+      0.95
+    ],
+    "optimizer_eps": 1e-08,
+    "lora_rank": 64,
+    "lora_alpha": 128,
+    "lora_dropout": 0.0,
+    "lora_target_modules": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "lora_weight_decay": 0.1,
+    "row_weight_decay": 0.0,
+    "max_grad_norm": 1.0,
+    "tokens_per_optimizer_update": 65536,
+    "packing": false,
+    "truncation": false,
+    "repeat_examples": false,
+    "precision": "bfloat16",
+    "trainable_dtype": "float32",
+    "gradient_checkpointing": true,
+    "eos_weight": 1.0,
+    "scheduler": "cosine_by_assistant_target_token_dose",
+    "worker_lr_scale": "rstar_dual_worker_v4._lr_scale",
+    "supervised_tail": [
+      151643
+    ],
+    "trainable_token_indices": {
+      "embed_tokens": [
+        151643,
+        151667,
+        151668
+      ],
+      "lm_head": [
+        151643,
+        151667,
+        151668
+      ]
+    },
+    "untie_tied_embeddings": true,
+    "context": 32768,
+    "eval": {
+      "suite": "DEV256",
+      "seed": 3407,
+      "temperature": 0.6,
+      "top_p": 0.95,
+      "top_k": 20,
+      "ctx": 32768,
+      "stop_ids": [
+        151643,
+        151645
+      ],
+      "base_reeval_same_mode": true
+    }
+  },
+  "code": {
+    "path": "/workspace/tools/rstar_dual_worker_v4.py",
+    "commit_or_sha256": "5cbdaca65556cbe0cf9e1374cf122c22511b02781eda6fab4ca780a869562cb7",
+    "prep": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/scripts/build_mix_distill_payload.py",
+    "prep_sha256": "2ff0f61e200d18dd55b1301f9573e923abd6fe87608220f7126cf709d4cb7288"
+  },
+  "started_at": "2026-09-09T06:13:03Z",
+  "endpoint": "2EP_ONLY @64848450 assistant tokens; no checkpoint picking",
+  "status": "RUNNING"
+}

provenance/TRAINING_CONFIG.json

@@ -0,0 +1,196 @@
+{
+  "schema": "code-sixarm-training-config/v4",
+  "status": "READY_NOT_STARTED",
+  "config_id": "Q4B-THINK-MIX",
+  "arm_id": "Q4B-THINK-MIX",
+  "corrects": "Q4B-THINK Mix Distillation 0.2:0.8 (30B:4B by rows)",
+  "base": {
+    "model_id": "Qwen/Qwen3-4B-Base",
+    "revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+    "local_path": "/workspace/code-sft-infra/models/qwen3-4b-base",
+    "pure_base": true,
+    "warm_start_adapter": null,
+    "tied_embeddings": true,
+    "tokenizer_sha256": "c0382117ea329cdf097041132f6d735924b697924d6f6fc3945713e96ce87539",
+    "tokenizer_config_sha256": "3c04ed3ca964ea2f6b2b5faf0dc4d31aec1cb1e8b4bcf63f402d295046b422b5"
+  },
+  "boundary": {
+    "mode": "trainable_token_rows_both_sides",
+    "token_ids": [
+      151643,
+      151667,
+      151668
+    ],
+    "token_names": {
+      "151643": "<|endoftext|>",
+      "151667": "<think>",
+      "151668": "</think>"
+    },
+    "row_lr": 0.0001,
+    "row_weight_decay": 0.0,
+    "fp32_master": true,
+    "untie_tied_embeddings": true
+  },
+  "context": {
+    "model_context": 32768,
+    "packing": false,
+    "truncation": false
+  },
+  "checkpoint": {
+    "milestone_updates": [
+      495,
+      744,
+      990
+    ],
+    "checkpoint_target_tokens": [
+      32445505,
+      48759220,
+      64848450
+    ],
+    "save_only_do_not_pick": true,
+    "right_side_3_only": true,
+    "primary": "2EP_FINAL_ONLY",
+    "note": "right-side: first update fully past 1.0ep, 1.5ep, plus 2ep endpoint"
+  },
+  "data": {
+    "path": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/mix_l0.2_s0.8_tail151643_2EP.parquet",
+    "sha256": "6097d7bf8bc18621ed4fbc2c7f4e6a68b65186eb1d25964fd632580f97dabdfd",
+    "unique_rows": 4155,
+    "payload_rows": 8310,
+    "physical_rows": 8310,
+    "epochs": 2,
+    "physical_2ep_rule": "concat(mix_l0.2_s0.8, mix_l0.2_s0.8) same order; 1ep is unique mixed rows",
+    "planned_active_tokens": 64848450,
+    "active_tokens_1ep": 32424225,
+    "full_tokens_1ep": 34594765,
+    "optimizer_updates": 990,
+    "census": "/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907/arms/Q4B-THINK-MIX/data/TOKEN_CENSUS_MIX.json",
+    "duplicate_rows_allowed": true,
+    "order_mutation_allowed": false,
+    "token_columns_in_parquet": "supervised tail [151643]; matches v4 worker render",
+    "source_1ep_sha256": "3c5ae286e7e5242df3c011123edaf0727405d3b535e85d77ce6b99afe7c27e77"
+  },
+  "model_runtime": {
+    "bf16": true,
+    "trainable_dtype": "float32",
+    "finetuning_type": "lora",
+    "lora_rank": 64,
+    "lora_alpha": 128,
+    "lora_dropout": 0.0,
+    "lora_targets": [
+      "q_proj",
+      "k_proj",
+      "v_proj",
+      "o_proj",
+      "gate_proj",
+      "up_proj",
+      "down_proj"
+    ],
+    "modules_to_save": null,
+    "gradient_checkpointing": true,
+    "trainable_token_indices": {
+      "embed_tokens": [
+        151643,
+        151667,
+        151668
+      ],
+      "lm_head": [
+        151643,
+        151667,
+        151668
+      ]
+    }
+  },
+  "optimization": {
+    "active_tokens_per_update_target": 65536,
+    "betas": [
+      0.9,
+      0.95
+    ],
+    "epsilon": 1e-08,
+    "learning_rate": 0.0001,
+    "max_grad_norm": 1.0,
+    "optimizer": "adamw",
+    "weight_decay": 0.1,
+    "whole_row_microsteps": true,
+    "warmup_fraction": 0.06,
+    "warmup_target_tokens": 3890907,
+    "scheduler": "cosine_by_assistant_target_token_dose",
+    "min_lr_ratio": 0.0,
+    "scheduler_horizon_optimizer_updates": 990,
+    "target_assistant_tokens": 64848450,
+    "worker_lr_scale": "rstar_dual_worker_v4._lr_scale"
+  },
+  "performance_recipe": {
+    "eos_weight": 1.0,
+    "prompt_mask": -100
+  },
+  "recipe_worker_fields": {
+    "tokens_per_optimizer_update": 65536,
+    "max_sequence_length": 32768,
+    "packing": {
+      "enabled": false
+    },
+    "repeat_examples": false,
+    "note": "2EP is physical concat; do not recycle via repeat_examples"
+  },
+  "reproducibility": {
+    "seed": 42
+  },
+  "runner": {
+    "path": "/workspace/tools/rstar_dual_worker_v4.py",
+    "sha256": "5cbdaca65556cbe0cf9e1374cf122c22511b02781eda6fab4ca780a869562cb7",
+    "rewrite_forbidden": true
+  },
+  "template": {
+    "id": "qwen3-8b-think-lcb-system-v2",
+    "mode": "think",
+    "enable_thinking": true,
+    "generation_prefill": false,
+    "target_contains_full_think": true,
+    "canonical_empty_think_pair_in_prompt": false,
+    "empty_think_pair_in_loss": false,
+    "target_forbids_think_tags": false,
+    "olmo_nothink_empty_pair_equivalent": null,
+    "chat_template_file": null,
+    "system_prompt_sha256": "41b9af0ebedb936e74e0996175f76f0116c334375ec3a7c914166f126c02f489",
+    "chat_template_sha256": "87a2728cb8dc9fe424d624542f6060ec05a1d285ebbec578bb078900e33396b5",
+    "target_tail": [
+      151643
+    ],
+    "evaluation_stop_ids": [
+      151643,
+      151645
+    ],
+    "eos_token_ids": [
+      151643
+    ],
+    "eos_token": "<|endoftext|>",
+    "supervision_eos": 151643,
+    "forbidden_label_ids": [
+      151645,
+      198
+    ]
+  },
+  "eval": {
+    "suite": "DEV256",
+    "seed": 3407,
+    "enable_thinking": true,
+    "canonical_empty_think_pair": false,
+    "generated_think_tags": null,
+    "temperature": 0.6,
+    "top_p": 0.95,
+    "top_k": 20,
+    "model_context": 32768,
+    "max_actual": "ctx-prompt-64",
+    "stop_ids": [
+      151643,
+      151645
+    ],
+    "base_must_reeval_same_mode": true,
+    "primary_checkpoint": "2EP_FINAL_ONLY",
+    "milestones": "appendix_only",
+    "cap_rate": "report_only",
+    "do_not_execute": true
+  }
+}

provenance/build_mix_distill_payload.py

@@ -0,0 +1,789 @@
+#!/usr/bin/env python3
+"""Build Mix Distillation 1ep/2ep payload for Q4B-THINK-MIX (Li et al. 2025 Mix-Large).
+
+30B-A3B-Thinking : 4B-Thinking = 0.2 : 0.8 by row/problem count. One teacher
+trace per problem. Dose target 32,436,894 assistant tokens/ep (±1%), physical
+2ep concat same order. V4 recipe otherwise unchanged.
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+import math
+import random
+import sys
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any
+
+import pyarrow as pa
+import pyarrow.parquet as pq
+
+ROOT = Path("/workspace/code-sft-runs/t30b2507-q4b-tn-v2-lr1e4-e1-20260907")
+ARM = ROOT / "arms" / "Q4B-THINK-MIX"
+DATA = ARM / "data"
+P30_2EP = ROOT / "data" / "THINK_PAIRED_V4_2EP.parquet"
+P30_2EP_SHA = "c160713fea76eae124dbbc3b5c619d3164862da83f61d91d3970ede990a12540"
+P30_1EP_ROWS = 4715
+P4 = ROOT / "arms" / "Q4B-THINK-T4BDATA" / "data" / "joint-full-t4b_tail151643.parquet"
+P4_SHA = "d09c4d6272de7e50c9b2df0e401878864f09bda33db05a5833c43575969fa524"
+P4_ROWS = 3866
+OUT_1EP = DATA / "mix_l0.2_s0.8_tail151643.parquet"
+OUT_2EP = DATA / "mix_l0.2_s0.8_tail151643_2EP.parquet"
+READY = DATA / "mix_l0.2_s0.8_tail151643_2EP_READY.json"
+CENSUS = DATA / "TOKEN_CENSUS_MIX.json"
+MQ0_OUT = DATA / "MQ0_NEAR_DUP.json"
+V4_POLICY = ROOT / "arms" / "Q4B-THINK" / "POLICY.json"
+V4_CFG = ROOT / "arms" / "Q4B-THINK-T4BDATA" / "TRAINING_CONFIG.json"
+V4_LAUNCH = ROOT / "arms" / "Q4B-THINK-T4BDATA" / "control" / "LAUNCH.json"
+WORKER = Path("/workspace/tools/rstar_dual_worker_v4.py")
+RUN_ID = "t30b2507-q4b-think-mix-l0.2-s0.8-v4-tail151643"
+ARM_ID = "Q4B-THINK-MIX"
+SEED = 42
+TARGET_1EP = 32_436_894
+TOL = 0.01
+RATIO_30, RATIO_4 = 1, 4
+
+sys.path.insert(0, str(ROOT / "scripts"))
+sys.path.insert(0, "/workspace/code-sft-infra/src")
+sys.path.insert(0, "/workspace/tools")
+sys.path.insert(0, "/workspace/code_think_30b_main_20260906/scripts")
+
+from code_sft_infra.contracts import document_digest, sha256_file  # noqa: E402
+from code_sft_infra.evaluation import build_lcb_prompt_messages  # noqa: E402
+from code_sft_infra.rendering import render_completion_only  # noqa: E402
+from code_sft_infra.source_census import load_hash_denylist  # noqa: E402
+import materialize_t30b_g2_fenced as think_m  # noqa: E402
+import near_dup_scan  # noqa: E402
+import select_core  # noqa: E402
+import v4_render_lib  # noqa: E402
+
+
+def _sha256_bytes_file(path: Path) -> str:
+    digest = hashlib.sha256()
+    with path.open("rb") as handle:
+        for chunk in iter(lambda: handle.read(8 << 20), b""):
+            digest.update(chunk)
+    return digest.hexdigest()
+
+
+def _pct(values: list[int], p: float) -> float:
+    if not values:
+        return float("nan")
+    xs = sorted(values)
+    if len(xs) == 1:
+        return float(xs[0])
+    k = (len(xs) - 1) * p / 100.0
+    lo = int(math.floor(k))
+    hi = int(math.ceil(k))
+    if lo == hi:
+        return float(xs[lo])
+    return xs[lo] + (xs[hi] - xs[lo]) * (k - lo)
+
+
+def mq0_all_hits(rows: list[dict[str, Any]]) -> tuple[dict[str, Any], list[dict[str, Any]]]:
+    summary = think_m.run_mq0(rows)
+    if summary.get("coverage") != "6_OF_6":
+        raise SystemExit(f"MQ0 coverage not 6_OF_6: {summary}")
+    if summary.get("status") in {"blocked_before_train"}:
+        raise SystemExit(f"MQ0 blocked: {summary}")
+    denylist, _ = load_hash_denylist(think_m.DENYLIST_PATH)
+    extra, _ = think_m.load_eval_extra_signatures()
+    union = set(denylist) | extra
+    hits: list[dict[str, Any]] = []
+    for row in rows:
+        overlap = sorted(think_m.train_signatures(row) & union)
+        if overlap:
+            hits.append(
+                {
+                    "problem_hash": row["problem_hash"],
+                    "problem_id": row.get("problem_id"),
+                    "overlap_n": len(overlap),
+                    "overlap_head": overlap[:3],
+                    "reason": "mq0_intersection",
+                }
+            )
+    if len(hits) != int(summary.get("intersection") or 0):
+        raise SystemExit(
+            f"MQ0 hit count mismatch full={len(hits)} summary={summary.get('intersection')}"
+        )
+    return summary, hits
+
+
+def _index(rows: list[dict[str, Any]], label: str) -> dict[str, dict[str, Any]]:
+    out: dict[str, dict[str, Any]] = {}
+    pids: dict[str, str] = {}
+    for row in rows:
+        h = str(row["problem_hash"])
+        pid = str(row["problem_id"])
+        if h in out:
+            raise SystemExit(f"{label} duplicate problem_hash {h}")
+        if pid in pids:
+            raise SystemExit(f"{label} duplicate problem_id {pid}")
+        out[h] = row
+        pids[pid] = h
+    return out
+
+
+def _fill(
+    union: list[str],
+    preferred: dict[str, str],
+    only30: set[str],
+    only4: set[str],
+    n30: int,
+    n4: int,
+) -> tuple[list[str], list[str]]:
+    picked_30: list[str] = []
+    picked_4: list[str] = []
+    for h in union:
+        if len(picked_30) >= n30 and len(picked_4) >= n4:
+            break
+        if h in only30:
+            if len(picked_30) < n30:
+                picked_30.append(h)
+            continue
+        if h in only4:
+            if len(picked_4) < n4:
+                picked_4.append(h)
+            continue
+        pref = preferred[h]
+        if pref == "30b" and len(picked_30) < n30:
+            picked_30.append(h)
+        elif pref == "4b" and len(picked_4) < n4:
+            picked_4.append(h)
+        elif pref == "30b" and len(picked_4) < n4:
+            picked_4.append(h)
+        elif pref == "4b" and len(picked_30) < n30:
+            picked_30.append(h)
+    return picked_30, picked_4
+
+
+def _sample_record(i: int, row: dict[str, Any], tok: Any) -> dict[str, Any]:
+    reason = str(row.get("reasoning") or "")
+    answer = str(row.get("final_answer") or "")
+    rendered = v4_render_lib.render_qwen(
+        tok,
+        mode="think",
+        prompt=str(row["prompt"] or ""),
+        reasoning=reason,
+        answer=answer,
+    )
+    rec = {
+        "i": i,
+        "teacher_source": row["teacher_source"],
+        "problem_id": row["problem_id"],
+        "problem_hash": row["problem_hash"],
+        "prompt_head": str(row["prompt"])[:160].replace("\n", " "),
+        "think_in_reason": "<think>" in reason,
+        "think_close_in_reason": "</think>" in reason,
+        "code_fence_in_answer": "```" in answer,
+        "final_answer_head": answer[:120].replace("\n", "\\n"),
+        "assistant_target_tokens": int(row["assistant_target_tokens"]),
+        "full_tokens": int(row["full_tokens"]),
+        "rendered_think_open": rendered.target_text.lower().count("<think>"),
+        "rendered_think_close": rendered.target_text.lower().count("</think>"),
+        "rendered_code_fence": "```" in rendered.target_text,
+        "last_token": rendered.last_token_id,
+        "active_match": rendered.active_tokens == int(row["assistant_target_tokens"]),
+        "full_match": rendered.full_tokens == int(row["full_tokens"]),
+    }
+    if rec["rendered_think_open"] != 1 or rec["rendered_think_close"] != 1:
+        raise SystemExit(f"think pair not 1/1 i={i} {rec}")
+    if not rec["rendered_code_fence"]:
+        raise SystemExit(f"missing code fence i={i} {rec}")
+    if rec["last_token"] != 151643:
+        raise SystemExit(f"tail not 151643 i={i} last={rec['last_token']}")
+    return rec
+
+
+def main() -> int:
+    DATA.mkdir(parents=True, exist_ok=True)
+    for sub in ("out", "status", "logs", "control"):
+        (ARM / sub).mkdir(parents=True, exist_ok=True)
+
+    sha30 = _sha256_bytes_file(P30_2EP)
+    if sha30 != P30_2EP_SHA:
+        raise SystemExit(f"30B 2EP sha mismatch {sha30} != {P30_2EP_SHA}")
+    sha4 = _sha256_bytes_file(P4)
+    if sha4 != P4_SHA:
+        raise SystemExit(f"4B 1ep sha mismatch {sha4} != {P4_SHA}")
+
+    t30_all = pq.read_table(P30_2EP)
+    if t30_all.num_rows != P30_1EP_ROWS * 2:
+        raise SystemExit(f"30B 2EP rows {t30_all.num_rows} != {P30_1EP_ROWS * 2}")
+    t30 = t30_all.slice(0, P30_1EP_ROWS)
+    t4 = pq.read_table(P4)
+    if t4.num_rows != P4_ROWS:
+        raise SystemExit(f"4B rows {t4.num_rows} != {P4_ROWS}")
+    if list(t30.schema.names) != list(t4.schema.names):
+        raise SystemExit(f"schema names differ 30B={t30.schema.names} 4B={t4.schema.names}")
+    for name in t30.schema.names:
+        if str(t30.schema.field(name).type) != str(t4.schema.field(name).type):
+            raise SystemExit(f"schema type differ {name}")
+
+    rows30 = t30.to_pylist()
+    rows4 = t4.to_pylist()
+    tok30 = sum(int(r["assistant_target_tokens"]) for r in rows30)
+    tok4_pool = sum(int(r["assistant_target_tokens"]) for r in rows4)
+    if tok30 != TARGET_1EP:
+        raise SystemExit(f"30B 1ep tokens {tok30} != {TARGET_1EP}")
+    if tok4_pool != 30_327_765:
+        raise SystemExit(f"4B 1ep tokens {tok4_pool} != 30327765")
+
+    map30 = _index(rows30, "30B")
+    map4 = _index(rows4, "4B")
+    h30, h4 = set(map30), set(map4)
+    inter = h30 & h4
+    only30 = h30 - h4
+    only4 = h4 - h30
+    union_set = h30 | h4
+    pid30 = {str(map30[h]["problem_id"]) for h in h30}
+    pid4 = {str(map4[h]["problem_id"]) for h in h4}
+    if len(pid30 & pid4) != len(inter):
+        raise SystemExit("problem_id intersection size != problem_hash intersection")
+    for h in inter:
+        if str(map30[h]["problem_id"]) != str(map4[h]["problem_id"]):
+            raise SystemExit(f"hash/id mismatch {h}")
+
+    print(
+        "UNIVERSE",
+        json.dumps(
+            {
+                "union": len(union_set),
+                "intersect": len(inter),
+                "only30": len(only30),
+                "only4": len(only4),
+                "pid_intersect": len(pid30 & pid4),
+                "pid_only30": len(pid30 - pid4),
+                "pid_only4": len(pid4 - pid30),
+            }
+        ),
+    )
+
+    mean30 = tok30 / len(rows30)
+    mean4 = tok4_pool / len(rows4)
+    n30_est = max(1, int(round(TARGET_1EP / (mean30 * RATIO_30 + mean4 * RATIO_4))))
+    n4_est = n30_est * RATIO_4
+    print("EST", n30_est, n4_est, "mean30", mean30, "mean4", mean4)
+    need4_tok = 0.8 * TARGET_1EP
+    shortfall_notes: list[str] = []
+    if n4_est > len(h4):
+        shortfall_notes.append(
+            f"4B pool rows {len(h4)} < needed {n4_est} for 1:4; not changing ratio"
+        )
+    if tok4_pool < need4_tok:
+        shortfall_notes.append(
+            f"4B pool tokens {tok4_pool} < 0.8*target {need4_tok}; not changing ratio"
+        )
+    if shortfall_notes:
+        print("SHORTFALL", shortfall_notes)
+
+    rng = random.Random(SEED)
+    union = sorted(union_set)
+    rng.shuffle(union)
+    preferred: dict[str, str] = {}
+    for h in union:
+        if h in inter:
+            preferred[h] = "30b" if rng.choice((0, 1)) == 0 else "4b"
+
+    def tokens_of(p30: list[str], p4: list[str]) -> int:
+        return sum(int(map30[h]["assistant_target_tokens"]) for h in p30) + sum(
+            int(map4[h]["assistant_target_tokens"]) for h in p4
+        )
+
+    candidates: list[tuple[int, int, list[str], list[str], int]] = []
+    n30_max = min(len(h30), len(h4) // RATIO_4)
+    for n30 in range(1, n30_max + 1):
+        n4 = n30 * RATIO_4
+        p30, p4 = _fill(union, preferred, only30, only4, n30, n4)
+        if len(p30) != n30 or len(p4) != n4:
+            continue
+        tok = tokens_of(p30, p4)
+        candidates.append((abs(tok - TARGET_1EP), n30, p30, p4, tok))
+
+    if not candidates:
+        raise SystemExit("no exact 1:4 fill succeeded")
+    in_band = [c for c in candidates if abs(c[4] - TARGET_1EP) <= TARGET_1EP * TOL]
+    pool = in_band if in_band else candidates
+    pool.sort(key=lambda c: (c[0], abs(c[1] - n30_est)))
+    _, n30, picked_30, picked_4, active_1ep = pool[0]
+    n4 = n30 * RATIO_4
+    rel = abs(active_1ep - TARGET_1EP) / TARGET_1EP
+    print(
+        "PICKED",
+        json.dumps(
+            {
+                "n30": n30,
+                "n4": n4,
+                "active_1ep": active_1ep,
+                "rel_err": rel,
+                "in_band": rel <= TOL,
+                "n30_est": n30_est,
+                "candidates_in_band": len(in_band),
+            }
+        ),
+    )
+    if rel > TOL:
+        raise SystemExit(f"token dose {active_1ep} outside ±1% of {TARGET_1EP}")
+
+    mixed: list[dict[str, Any]] = []
+    for h in picked_30:
+        row = dict(map30[h])
+        row["teacher_source"] = "30b"
+        mixed.append(row)
+    for h in picked_4:
+        row = dict(map4[h])
+        row["teacher_source"] = "4b"
+        mixed.append(row)
+    if len({r["problem_hash"] for r in mixed}) != len(mixed):
+        raise SystemExit("duplicate problem in mix")
+    order_rng = random.Random(SEED)
+    order_rng.shuffle(mixed)
+
+    mq0_summary, mq0_hits = mq0_all_hits(mixed)
+    print(
+        "mq0",
+        mq0_summary.get("status"),
+        "intersection",
+        mq0_summary.get("intersection"),
+        "coverage",
+        mq0_summary.get("coverage"),
+    )
+    if mq0_hits or int(mq0_summary.get("intersection") or 0) != 0:
+        raise SystemExit(f"MQ0 intersection nonzero: {mq0_hits[:5]}")
+    if mq0_summary.get("coverage") != "6_OF_6":
+        raise SystemExit(f"MQ0 coverage {mq0_summary.get('coverage')}")
+
+    near = near_dup_scan.scan_training_rows(mixed)
+    print("near_dup hits", len(near["hit_hashes"]), "grey", near["grey_zone_count"])
+    if near["hit_hashes"]:
+        raise SystemExit(f"near-dup hits: {near['hit_hashes']}")
+
+    tok = v4_render_lib.load_qwen_tokenizer()
+    idx30 = [i for i, r in enumerate(mixed) if r["teacher_source"] == "30b"][:3]
+    idx4 = [i for i, r in enumerate(mixed) if r["teacher_source"] == "4b"][:3]
+    samples = [_sample_record(i, mixed[i], tok) for i in idx30 + idx4]
+    if len(samples) != 6:
+        raise SystemExit(f"expected 6 samples, got {len(samples)}")
+
+    v4_policy = json.loads(V4_POLICY.read_text())
+    renderer = v4_policy["arms"][0]["renderer"]
+    authority = v4_policy["arms"][0]["renderer_authority"]
+    contract = {"renderer": renderer, "sequence": {"max_sequence_length": 32768}}
+    for rec in samples:
+        row = mixed[rec["i"]]
+        worker = render_completion_only(
+            contract,
+            renderer_id="qwen3-8b-think-lcb-system-v2",
+            prompt_messages=build_lcb_prompt_messages(contract, row["prompt"]),
+            reasoning=row["reasoning"],
+            answer=row["final_answer"],
+        )
+        if worker.labels[-1] != 151643:
+            raise SystemExit(f"worker last label {worker.labels[-1]} i={rec['i']}")
+        rec["worker_last"] = worker.labels[-1]
+        rec["worker_active"] = sum(x != -100 for x in worker.labels)
+        rec["worker_full"] = len(worker.input_ids)
+        print("SAMPLE", json.dumps(rec, ensure_ascii=False))
+
+    schema = t30.schema
+    arrays = []
+    for name in schema.names:
+        arrays.append(pa.array([r[name] for r in mixed], type=schema.field(name).type))
+    arrays.append(pa.array([r["teacher_source"] for r in mixed], type=pa.string()))
+    names = list(schema.names) + ["teacher_source"]
+    one = pa.table(arrays, names=names)
+    two = pa.concat_tables([one, one])
+    if OUT_1EP.exists() or OUT_2EP.exists():
+        raise SystemExit("refusing to overwrite existing mix parquet")
+    pq.write_table(one, OUT_1EP)
+    pq.write_table(two, OUT_2EP)
+    sha1 = _sha256_bytes_file(OUT_1EP)
+    sha2 = _sha256_bytes_file(OUT_2EP)
+    actives = [int(r["assistant_target_tokens"]) for r in mixed]
+    fulls = [int(r["full_tokens"]) for r in mixed]
+    actives30 = [int(r["assistant_target_tokens"]) for r in mixed if r["teacher_source"] == "30b"]
+    actives4 = [int(r["assistant_target_tokens"]) for r in mixed if r["teacher_source"] == "4b"]
+    active_2ep = active_1ep * 2
+    print("wrote", OUT_1EP, one.num_rows, sha1, "active1ep", active_1ep)
+    print("wrote", OUT_2EP, two.num_rows, sha2, "active2ep", active_2ep)
+
+    stream_rows = [
+        {"problem_hash": r["problem_hash"], "active_tokens": int(r["assistant_target_tokens"])}
+        for r in mixed
+    ]
+    derived = select_core.derive_after_materialize(stream_rows)
+    if derived["planned_active_tokens_2ep"] != active_2ep:
+        raise SystemExit("derived 2ep tokens != 2*1ep")
+    print(
+        "derived",
+        json.dumps(
+            {
+                k: derived[k]
+                for k in (
+                    "unique_rows",
+                    "physical_rows_2ep",
+                    "unique_1ep_think_active_tokens",
+                    "planned_active_tokens_2ep",
+                    "optimizer_updates",
+                    "warmup_target_tokens",
+                    "milestone_updates_right_3",
+                    "checkpoint_target_tokens_right_3",
+                )
+            }
+        ),
+    )
+
+    mq0_body = {
+        "mq0": {
+            "status": mq0_summary.get("status"),
+            "coverage": mq0_summary.get("coverage"),
+            "intersection": mq0_summary.get("intersection"),
+            "denylist_path": mq0_summary.get("denylist_path"),
+            "denylist_sha256": mq0_summary.get("denylist_sha256"),
+        },
+        "near_dup_hit_hashes": near["hit_hashes"],
+        "near_dup_grey_zone_count": near["grey_zone_count"],
+        "dropped_n": 0,
+        "kept_n": len(mixed),
+    }
+    MQ0_OUT.write_text(json.dumps(mq0_body, indent=2, ensure_ascii=False) + "\n")
+
+    census = {
+        "schema": "TOKEN_CENSUS_MIX_L0.2_S0.8_TAIL151643",
+        "seed": SEED,
+        "n_rows": len(mixed),
+        "n_problems": len(mixed),
+        "parquet_1ep": str(OUT_1EP),
+        "parquet_1ep_sha256": sha1,
+        "parquet_2ep": str(OUT_2EP),
+        "parquet_2ep_sha256": sha2,
+        "supervised_tail": [151643],
+        "active_1ep": active_1ep,
+        "active_2ep": active_2ep,
+        "full_1ep": int(sum(fulls)),
+        "target_1ep": TARGET_1EP,
+        "rel_err": rel,
+        "ratio_rows_30_to_4": f"{n30}:{n4}",
+        "universe": {
+            "union": len(union_set),
+            "intersect": len(inter),
+            "only30": len(only30),
+            "only4": len(only4),
+        },
+        "teacher": {
+            "n30": n30,
+            "n4": n4,
+            "tok30": int(sum(actives30)),
+            "tok4": int(sum(actives4)),
+            "row_frac_30": n30 / len(mixed),
+            "row_frac_4": n4 / len(mixed),
+            "tok_frac_30": sum(actives30) / active_1ep,
+            "tok_frac_4": sum(actives4) / active_1ep,
+            "n30_from_intersect": sum(1 for h in picked_30 if h in inter),
+            "n30_from_only30": sum(1 for h in picked_30 if h in only30),
+            "n4_from_intersect": sum(1 for h in picked_4 if h in inter),
+            "n4_from_only4": sum(1 for h in picked_4 if h in only4),
+        },
+        "shortfall_notes": shortfall_notes,
+        "active": {
+            "n": len(actives),
+            "sum": active_1ep,
+            "min": min(actives),
+            "p10": _pct(actives, 10),
+            "p50": _pct(actives, 50),
+            "p90": _pct(actives, 90),
+            "p95": _pct(actives, 95),
+            "max": max(actives),
+        },
+        "active_30b": {
+            "n": len(actives30),
+            "sum": int(sum(actives30)),
+            "min": min(actives30),
+            "p50": _pct(actives30, 50),
+            "max": max(actives30),
+        },
+        "active_4b": {
+            "n": len(actives4),
+            "sum": int(sum(actives4)),
+            "min": min(actives4),
+            "p50": _pct(actives4, 50),
+            "max": max(actives4),
+        },
+        "derived": {
+            k: derived[k]
+            for k in (
+                "unique_rows",
+                "physical_rows_2ep",
+                "unique_1ep_think_active_tokens",
+                "planned_active_tokens_2ep",
+                "optimizer_updates",
+                "warmup_target_tokens",
+                "milestone_updates_right_3",
+                "checkpoint_target_tokens_right_3",
+            )
+        },
+        "samples": samples,
+        "sources": {
+            "30b_2ep": str(P30_2EP),
+            "30b_2ep_sha256": sha30,
+            "30b_1ep_rows_used": P30_1EP_ROWS,
+            "4b_1ep": str(P4),
+            "4b_1ep_sha256": sha4,
+        },
+    }
+    CENSUS.write_text(json.dumps(census, indent=2, ensure_ascii=False) + "\n")
+
+    ready_body = {
+        "physical_2ep": "concat(mix_table, mix_table) same order; tail [151643]; teacher_source kept",
+        "schema": "CODE_SFT_OR1RECIPE_READY_V1",
+        "status": "READY",
+        "training_payload": {
+            "active_tokens": active_2ep,
+            "path": OUT_2EP.name,
+            "rows": two.num_rows,
+            "sha256": sha2,
+        },
+    }
+    READY.write_text(json.dumps(ready_body, indent=2, sort_keys=True) + "\n")
+    ready_sha = sha256_file(READY)
+
+    cfg = json.loads(V4_CFG.read_text())
+    cfg["status"] = "READY_NOT_STARTED"
+    cfg["config_id"] = ARM_ID
+    cfg["arm_id"] = ARM_ID
+    cfg["corrects"] = "Q4B-THINK Mix Distillation 0.2:0.8 (30B:4B by rows)"
+    cfg["checkpoint"]["milestone_updates"] = derived["milestone_updates_right_3"]
+    cfg["checkpoint"]["checkpoint_target_tokens"] = derived["checkpoint_target_tokens_right_3"]
+    cfg["data"]["path"] = str(OUT_2EP)
+    cfg["data"]["sha256"] = sha2
+    cfg["data"]["unique_rows"] = len(mixed)
+    cfg["data"]["payload_rows"] = two.num_rows
+    cfg["data"]["physical_rows"] = two.num_rows
+    cfg["data"]["physical_2ep_rule"] = (
+        "concat(mix_l0.2_s0.8, mix_l0.2_s0.8) same order; 1ep is unique mixed rows"
+    )
+    cfg["data"]["planned_active_tokens"] = active_2ep
+    cfg["data"]["active_tokens_1ep"] = active_1ep
+    cfg["data"]["full_tokens_1ep"] = int(sum(fulls))
+    cfg["data"]["optimizer_updates"] = derived["optimizer_updates"]
+    cfg["data"]["census"] = str(CENSUS)
+    cfg["data"]["token_columns_in_parquet"] = "supervised tail [151643]; matches v4 worker render"
+    cfg["data"]["source_1ep_sha256"] = sha1
+    cfg["data"].pop("source_old_payload", None)
+    cfg["data"].pop("source_old_sha256", None)
+    cfg["optimization"]["warmup_target_tokens"] = derived["warmup_target_tokens"]
+    cfg["optimization"]["scheduler_horizon_optimizer_updates"] = derived["optimizer_updates"]
+    cfg["optimization"]["target_assistant_tokens"] = active_2ep
+    worker_sha = _sha256_bytes_file(WORKER)
+    cfg["runner"]["sha256"] = worker_sha
+    (ARM / "TRAINING_CONFIG.json").write_text(json.dumps(cfg, indent=2, ensure_ascii=False) + "\n")
+
+    (ARM / "TOKEN_ROWS_META.json").write_text(
+        json.dumps(
+            {
+                "schema": "TOKEN_ROWS_META_V1",
+                "ids": [151643, 151667, 151668],
+                "dtype": "float32",
+                "tied_before": True,
+                "untied": True,
+                "worker": "rstar_dual_worker_v4",
+                "hidden": 2560,
+                "base_rev": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+            },
+            indent=2,
+        )
+        + "\n"
+    )
+
+    identity = {
+        "schema": "RUN_IDENTITY_V1",
+        "run_id": RUN_ID,
+        "corrects": "Q4B-THINK Mix Distillation 0.2:0.8 (30B-A3B-Thinking : 4B-Thinking by rows)",
+        "notes": (
+            "Mix Distillation Mix-Large (Li et al. 2025 arXiv 2502.12143): "
+            "30B-A3B-Thinking : 4B-Thinking = 0.2 : 0.8 by row/problem; one teacher "
+            f"trace per problem; dose target {TARGET_1EP}/ep ±1%, actual 1ep "
+            f"{active_1ep}, 2ep {active_2ep}. 30B source={P30_2EP} sha={sha30} "
+            f"(2EP first half {P30_1EP_ROWS} rows). 4B source={P4} sha={sha4}. "
+            "V4 recipe unchanged (LoRA r64/α128, lr 1e-4, both-sides B-row "
+            "[151643,151667,151668], physical 2ep concat, seed 42). "
+            "PI approved 2026-09-09."
+        ),
+        "model": {
+            "hf_id": "Qwen/Qwen3-4B-Base",
+            "revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+            "local_path": "/workspace/code-sft-infra/models/qwen3-4b-base",
+            "tied_embeddings": True,
+        },
+        "data": {
+            "path": str(OUT_2EP),
+            "rows": two.num_rows,
+            "sha256": sha2,
+            "census": str(CENSUS),
+            "physical_2ep": "concat(mix, mix) same order; teacher_source 30b/4b; tail 151643",
+            "source_1ep_path": str(OUT_1EP),
+            "source_1ep_sha256": sha1,
+            "source_1ep_rows": one.num_rows,
+            "teacher_30b_path": str(P30_2EP),
+            "teacher_30b_sha256": sha30,
+            "teacher_4b_path": str(P4),
+            "teacher_4b_sha256": sha4,
+            "mix_ratio_rows": "0.2:0.8",
+            "n30": n30,
+            "n4": n4,
+            "active_tokens_1ep": active_1ep,
+        },
+        "seeds": {"train": 42, "eval": 3407, "mix": 42},
+        "parameters": {
+            "epochs": 2,
+            "target_assistant_tokens": active_2ep,
+            "active_tokens_1ep": active_1ep,
+            "checkpoint_target_tokens": derived["checkpoint_target_tokens_right_3"],
+            "milestone_updates": derived["milestone_updates_right_3"],
+            "warmup_target_tokens": derived["warmup_target_tokens"],
+            "learning_rate": 0.0001,
+            "row_lr": 0.0001,
+            "optimizer": "adamw",
+            "optimizer_betas": [0.9, 0.95],
+            "optimizer_eps": 1e-08,
+            "lora_rank": 64,
+            "lora_alpha": 128,
+            "lora_dropout": 0.0,
+            "lora_target_modules": [
+                "q_proj",
+                "k_proj",
+                "v_proj",
+                "o_proj",
+                "gate_proj",
+                "up_proj",
+                "down_proj",
+            ],
+            "lora_weight_decay": 0.1,
+            "row_weight_decay": 0.0,
+            "max_grad_norm": 1.0,
+            "tokens_per_optimizer_update": 65536,
+            "packing": False,
+            "truncation": False,
+            "repeat_examples": False,
+            "precision": "bfloat16",
+            "trainable_dtype": "float32",
+            "gradient_checkpointing": True,
+            "eos_weight": 1.0,
+            "scheduler": "cosine_by_assistant_target_token_dose",
+            "worker_lr_scale": "rstar_dual_worker_v4._lr_scale",
+            "supervised_tail": [151643],
+            "trainable_token_indices": {
+                "embed_tokens": [151643, 151667, 151668],
+                "lm_head": [151643, 151667, 151668],
+            },
+            "untie_tied_embeddings": True,
+            "context": 32768,
+            "eval": {
+                "suite": "DEV256",
+                "seed": 3407,
+                "temperature": 0.6,
+                "top_p": 0.95,
+                "top_k": 20,
+                "ctx": 32768,
+                "stop_ids": [151643, 151645],
+                "base_reeval_same_mode": True,
+            },
+        },
+        "code": {
+            "path": str(WORKER),
+            "commit_or_sha256": worker_sha,
+            "prep": str(Path(__file__).resolve()),
+            "prep_sha256": _sha256_bytes_file(Path(__file__).resolve()),
+        },
+        "started_at": None,
+        "endpoint": f"2EP_ONLY @{active_2ep} assistant tokens; no checkpoint picking",
+        "status": "READY_NOT_STARTED",
+    }
+    (ARM / "RUN_IDENTITY.json").write_text(json.dumps(identity, indent=2, ensure_ascii=False) + "\n")
+
+    v4_launch = json.loads(V4_LAUNCH.read_text())
+    policy = {
+        "schema": "CODE_SFT_RSTAR_DUAL_TRAINING_POLICY_V1",
+        "status": "SEALED",
+        "owner_authorization": "PI 2026-09-09 Q4B-THINK-MIX Mix Distillation 0.2:0.8",
+        "generation_behavior_gate": None,
+        "arms": [
+            {
+                "arm_id": ARM_ID,
+                "gpu": 2,
+                "model_id": "qwen3-4b-base",
+                "source_id": "rstar-coder",
+                "renderer_id": "qwen3-8b-think-lcb-system-v2",
+                "renderer": renderer,
+                "renderer_authority": authority,
+                "model": json.loads(json.dumps(v4_policy["arms"][0]["model"])),
+                "data": {
+                    "assistant_target_tokens": active_2ep,
+                    "manifest_path": str(READY),
+                    "manifest_file_sha256": ready_sha,
+                    "manifest_sha256": ready_sha,
+                    "payload_rows": two.num_rows,
+                    "payload_sha256": sha2,
+                    "source_revision": "3a7a0a0636ec96e3c1ec42ebe79ade467caa040d",
+                    "max_assistant_target_tokens_per_example": 16384,
+                },
+            }
+        ],
+        "recipe": json.loads(json.dumps(v4_policy["recipe"])),
+        "runtime": {
+            "state_root": str(ARM),
+            "checkpoint_root": str(ARM / "out"),
+        },
+    }
+    policy["recipe"]["target_assistant_tokens"] = active_2ep
+    policy["recipe"]["warmup_target_tokens"] = derived["warmup_target_tokens"]
+    policy["recipe"]["checkpoint_target_tokens"] = derived["checkpoint_target_tokens_right_3"]
+    policy_path = ARM / "POLICY.json"
+    policy_path.write_text(json.dumps(policy, indent=2, sort_keys=True) + "\n")
+    policy_sha = sha256_file(policy_path)
+
+    job = {
+        "arm_id": ARM_ID,
+        "assistant_target_tokens": active_2ep,
+        "gpu": 2,
+        "materialized_manifest_path": str(READY),
+        "materialized_manifest_sha256": ready_sha,
+        "model_hf_id": "Qwen/Qwen3-4B-Base",
+        "model_revision": "906bfd4b4dc7f14ee4320094d8b41684abff8539",
+        "payload_sha256": sha2,
+        "renderer_authority_sha256": authority["sha256"],
+        "renderer_id": "qwen3-8b-think-lcb-system-v2",
+        "source_id": "rstar-coder",
+    }
+    job["job_sha256"] = document_digest(job, "job_sha256")
+    launch = {
+        "schema": "CODE_SFT_RSTAR_DUAL_LAUNCH_V1",
+        "status": "ATOMICALLY_RELEASED",
+        "run_id": RUN_ID,
+        "automation_policy_sha256": v4_launch["automation_policy_sha256"],
+        "delegation_sha256": v4_launch["delegation_sha256"],
+        "launch_critical_go_sha256": v4_launch["launch_critical_go_sha256"],
+        "release_sha256": v4_launch["release_sha256"],
+        "release_nonce": v4_launch["release_nonce"],
+        "release_claim_path": v4_launch["release_claim_path"],
+        "release_claim_file_sha256": v4_launch["release_claim_file_sha256"],
+        "evidence_path": v4_launch["evidence_path"],
+        "evidence_file_sha256": v4_launch["evidence_file_sha256"],
+        "training_policy_path": str(policy_path),
+        "training_policy_sha256": policy_sha,
+        "jobs": [job],
+    }
+    launch["launch_sha256"] = document_digest(launch, "launch_sha256")
+    launch_path = ARM / "control" / "LAUNCH.json"
+    launch_path.write_text(json.dumps(launch, indent=2, sort_keys=True) + "\n")
+    if document_digest(json.loads(launch_path.read_text()), "launch_sha256") != launch["launch_sha256"]:
+        raise SystemExit("launch digest drifted after write")
+    print("wrote", policy_path, policy_sha)
+    print("wrote", launch_path, launch["launch_sha256"])
+    print("PREP_OK", datetime.now(timezone.utc).isoformat())
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
+size 11422650

tokenizer_config.json

@@ -0,0 +1,28 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|endoftext|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}