SkyPanther/synack-decide-decision-index

SynACK Decide v7: Decision Index 0.2.1 run

A complete Decision Index 0.2.1 run of SkyPanther/synack-decide-26b-a4b (bf16). Decision Index 59.54, raw index 69.25.

Area	Skill
Knowledge & Reasoning	44.4
Language Understanding	63.9
Retrieval & Classification	62.4
Tools & Automation	79.1
Arts & Human Taste	45.9

runs/synack-decide-v7/:

• Scores: scores.json, index.json and benchmark-summary.json, exactly as the kit's pipeline wrote them (kit 87d4650, edition 0.2.1).
• Run record: environment.json and status.json.
• results.compact.jsonl.gz: every results row with payload and raw_output removed, the same fields the kit's --compact drops, so no suite text is published. Every row keeps its payload_sha256. Re-scoring it with python -m decision_index score --edition 0.2.1 at 87d4650 gives 59.54 again, with the same area scores.
• harness/job_full_eval.sh: the Hugging Face job that ran it.

How it was run

• One process, the kit's own runner: a single pipeline run, started fresh with no resume, on one NVIDIA RTX PRO 6000 Blackwell (96 GB). PyTorch 2.14.1+cu130 and transformers 5.18.0. It took 7 h 45 min.
• Weights: model commit c20b9b19. The repo history was later squashed into ec95c3f4, which has the same weight files.
• Engine: engine_v2:V2Engine. The exact code is in syn-ack-ai/synack-decide leaderboard/. Each question is one prompt, read in one forward pass. The engine takes a softmax over the option-label logits, so nothing is generated.
• Requests: all 150,317 scoreable requests returned ok. Of the 442 rows outside the 0.2.1 edition, 167 were refused as unsupported: ToolRet retrieval prompts of 34,000 to 38,444 tokens, over the engine's 32,768-token context limit. Those rows are never scored. Nothing was truncated, and no options were removed.
• Latency: median 131.7 ms and p95 360.2 ms per request, measured in-process, one request at a time and device-synchronized.

The suite itself is evaluation data under its upstream terms and is not included here.