Hopper (G) 1.3: new adapter weights (continued from 1.2); calibration map and serving unchanged; checksums and card

CHECKSUMS.txt

@@ -1,3 +1,3 @@
-36d1214ea16648b4b7cf690c779ddd84e7a089589a7a62652a1c01332c24d3e1  adapter_config.json
-8813cb2a1e44184bcdae017ace49cfc7cb00738c611423b6198c4b5ea553deaf  adapter_model.safetensors
+40d70c507c1db76454ac0db030fed751105d9b08148a24fdd9946b4252f26e30  adapter_config.json
+2737fdc5c282393a84cbc4f38baf92de5ec2031a9375eeedf891818871613051  adapter_model.safetensors
 217a2af396320c8a279e2f915fe17e67184cdf577668eb38147775184c259ade  hopper.json

README.md

@@ -10,57 +10,52 @@ tags: [decision, classification, calibration, lora]
 
 A general-purpose version of [Hopper](https://huggingface.co/HopitAI/hopper): a LoRA adapter for Qwen3.5-4B that
 answers typed decision questions in one forward pass by reading the probability of each option letter, with a
-per-kind calibration map. Served with the Hopper code at https://github.com/hopit-ai/hopper (tag `g-1.2.0`).
+per-kind calibration map. Served with the Hopper code at https://github.com/hopit-ai/hopper.
 
-**Research and demo use only.** This adapter continues training from Hopper 1.0's adapter, whose training data
-included passages from RACE (non-commercial research only), and its training data also includes material made with
-LLM-based generation. Do not use it commercially.
+**Current version: Hopper (G) 1.3** (tag `g-1.3.0`). Hopper (G) 1.2 stays available at
+revision `d60a1d6` and is what the Decision Index row below measures.
 
-## What it is
-
-- Base: Qwen/Qwen3.5-4B (revision `851bf6e`), LoRA rank 16, alpha 32, the same 12 modules as Hopper.
-- Training: continued from Hopper 1.0's adapter on Hopper's decision tasks (at maintenance doses), general-purpose
-  sources (tabular record joins, CLINC150 intents, GSM8K arithmetic) and replay of public training data, with a fixed
-  retention constraint against Hopper 1.0 on a held-out replay bank.
-- Serving: identical to Hopper 1.1.1, including the calibration map and the long-menu shortlist (more than 26
-  options answered in two disclosed stages).
+**Research and demo use only.** These adapters continue training from Hopper 1.0's adapter, whose training data
+included passages from RACE (non-commercial research only), and their training data also includes material made with
+LLM-based generation. Do not use them commercially.
 
 ## Leaderboards (official)
 
 - **[Jev Decision Index](https://huggingface.co/spaces/multimodalart/jev-decision-index)** (edition 0.2.1,
-  27 Sep 2026): **40.77, #16 of 68**, the highest of the 4B models (Decider 4B: 40.70). The row comes from a
-  complete run of the suite that we scored ourselves with the Index kit, at the maintainer's request. The model
-  outputs and scores are public at
-  [`HopitAI/hopper-g-decision-index-results`](https://huggingface.co/datasets/HopitAI/hopper-g-decision-index-results).
-  It replaced the Hopper 1.1.1 row (39.67).
-- **[JevBench](https://benchmarkheaven.com/jev-models)**: requested as a separate row
-  ([issue #112](https://github.com/fstandhartinger/jevbench/issues/112)), not yet measured. Our own development check
-  on held-out JevBench-style items (not an official score) put it level with Hopper 1.0 (+0.7 points, within noise),
-  so we make no JevBench improvement claim. Hopper 1.0's official JevBench result is 59.43, #6 of 90 ranked
-  (v1.4.2.1, 27 Sep 2026).
-
-## Evaluation (our runs)
-
-On our local run of the Decision Index 0.2 suite (40 benchmarks, A10G, same serving code, only the adapter
-differing):
-
-| | Hopper 1.1.1 | Hopper (G) 1.2 |
-| --- | ---: | ---: |
-| balanced raw | 52.74 | 53.50 |
-| balanced skill | 37.10 | 38.07 |
-| GSM8K | 0.318 | 0.480 |
-
-Paired bootstrap of the balanced-raw difference: +0.76 (95 % interval +0.55 to +0.98). Seed 1 (trained independently) confirms: balanced raw 53.44 vs 52.74 (+0.70), with every Index area at or above Hopper 1.1.1.
-Not official scores. Hopper (G) is not tuned for JevBench; we make no claim there.
-
-## Revisions
-
-- `d60a1d6`: the evaluated release (the revision given to the Decision Index and JevBench).
-- Later commit: `adapter_config.json` sets `"task_type": "CAUSAL_LM"` (it was `null`), for the Hub's metadata
-  parser only. The weights, calibration map and outputs are unchanged.
+  27 Sep 2026): Hopper (G) 1.2 scores **40.77, #16 of 68**, the highest of the 4B models (Decider 4B: 40.70), from a
+  complete self-scored run ([results](https://huggingface.co/datasets/HopitAI/hopper-g-decision-index-results)).
+  Hopper (G) 1.3 has not been scored on the Index yet.
+- **[JevBench](https://benchmarkheaven.com/jev-models)**: not yet measured. We have asked for Hopper (G) 1.3 to be
+  measured as a separate row ([issue #112](https://github.com/fstandhartinger/jevbench/issues/112)). Hopper 1.0's
+  official result is 59.43 (v1.4.2.2).
+
+## What changed in 1.3
+
+- **Weights only.** Continued from Hopper (G) 1.2 on 6,400 new code-generated decision items (dated arithmetic, long
+  policies with exceptions and precedence, multi-table lookups, ambiguity, probability, safety and judging checklists,
+  trade-offs, paraphrase pairs, injected-instruction traps), every answer computed and checked by code, plus
+  maintenance and replay of earlier training data under the same retention constraint as 1.2.
+- **Serving: unchanged.** Same code as Hopper 1.1.1 / Hopper (G) 1.2, same calibration map (byte-identical), eager
+  by default.
+
+## Evaluation (our runs, not official scores)
+
+On our private held-out decision set (2,700 items, nine decision families, built for this purpose and never trained
+on), 1.3 scores **+1.7 points over 1.2** (template bootstrap 95% interval +0.5 to +2.9). **This is below the +3.0 we
+pre-registered as our own bar for this build**, and a smaller run with a quarter of the new data did about as well
+(+2.1), so more of the same data did not help. We release it anyway as a disclosed, qualified release: on our
+regression checks against Hopper 1.0 (document reading, numeric, paraphrase, abstention, routing, calibration and a
+general-knowledge check) it passes every registered floor, with a pooled gain of +2.3 points (one-sided 95% lower
+bound +1.9) and hard-tier calibration 83.5 under the shipped map. None of this predicts a JevBench result.
 
 ## Limitations
 
 - English, 4B parameters; it reads options, it does not generate reasoning.
-- Calibration uses Hopper 1.1.1's map, fitted for Hopper 1.0's adapter; it has not been refitted for these weights.
+- It rarely concludes "no match" or "cannot be determined" when an explicitly incomplete record leaves a multi-step
+  lookup unresolved; 1.3 did not fix this.
 - Research and demo use only (see above).
+
+## Revisions
+
+- Tag `g-1.3.0`: Hopper (G) 1.3, the evaluated adapter bytes.
+- `d60a1d6`: Hopper (G) 1.2 (the Decision Index row). `060b1bd`: its config-only fix.

adapter_config.json

@@ -35,25 +35,25 @@
   "rank_pattern": {},
   "revision": null,
   "target_modules": [
+    "v_proj",
+    "down_proj",
+    "out_proj",
+    "in_proj_z",
     "in_proj_qkv",
     "q_proj",
-    "gate_proj",
-    "in_proj_z",
+    "k_proj",
     "in_proj_a",
-    "o_proj",
-    "v_proj",
-    "down_proj",
     "up_proj",
-    "k_proj",
-    "out_proj",
-    "in_proj_b"
+    "o_proj",
+    "in_proj_b",
+    "gate_proj"
   ],
   "target_parameters": null,
-  "task_type": "CAUSAL_LM",
+  "task_type": null,
   "trainable_token_indices": null,
   "use_bdlora": null,
   "use_dora": false,
   "use_qalora": false,
   "use_rslora": false,
   "velora_config": null
-}
+}

adapter_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8813cb2a1e44184bcdae017ace49cfc7cb00738c611423b6198c4b5ea553deaf
+oid sha256:2737fdc5c282393a84cbc4f38baf92de5ec2031a9375eeedf891818871613051
 size 129927008