code-nothink-q8b-20260908
Formerly Qwen3-8B-Base-code-v4-nothink (HF redirect).
Arm ID: Q8B-NOTHINK
Run ID: t30b2507-q8b-nothink-v4-tail151643
Merged full bf16 weights used for the official DEV256 NoThink evaluation. Tokenizer files in this repo are the eval-caliber bundle (qwen3-v4-tail151643); they overlay any tokenizer files that were present in the merge directory (the merge tree's tokenizer.json was a different, larger file).
Exploratory result (2 eval seeds), not a preregistered confirmatory claim.
Base model
- Hugging Face:
Qwen/Qwen3-8B-Base - Revision:
49e3418fbbbca6ecbdf9608b4d22e5a407081db4
Training
- Method: LoRA r64 / α128 on seven projections (
q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj), then merged into full-model bf16 safetensors - Data: NoThink code SFT (paired V4, physical 2-epoch concat)
- Endpoint (score): step 142, 9,455,736 assistant tokens — endpoint-as-score, no checkpoint picking
- Train seed 42; LR
1e-4; context 8192; AdamW; cosine by assistant-token dose - Host: local GPU 3 (retrain after mathbox3 host loss); eval tokenizer renderer
qwen3-v4-tail151643
Evaluation
- Suite: official LiveCodeBench DEV256
- Mode NoThink,
max_model_len8192, vLLM 0.28.0; do not prefill<think> - Metric: sandbox pass@1 = passed / 256
- Seed 12345 used the same merged weights, tokenizer, and vLLM command as seed 3407 except
--seedand the serve port. Results:/workspace/code-sft-runs/rstar-eval/official-dev256-seeds/seed12345/
| seed | arm | arm cap | Base | Base cap | McNemar arm_only / base_only | p |
|---|---|---|---|---|---|---|
| 3407 | 70/256 (27.3%) | 78 | 59/256 (23.0%) | 21 | 25 / 14 | 0.108 |
| 12345 | 84/256 (32.8%) | 75 | 57/256 (22.3%) | 23 | 37 / 10 | 9.8e-05 |
| pooled 512 | — | — | — | — | 62 / 24 | 5.1e-05 |
Mean: arm 77.0 (30.1%) vs Base 58.0 (22.7%), Δ ≈ +19.
Arm pass@1 moves a lot across seeds (70 vs 84); Base is stable (57–59). The two-seed pooled paired difference is significant (p = 5.1e-05). Still exploratory (2 seeds, not preregistered).
Provenance note
The same identity (seed 42, 9,455,736 assistant tokens, lr 1e-4, LoRA r64/α128) was previously trained once on another machine and scored 76/256 vs Base 57/256 (McNemar p=0.0019). That host went down and the weights plus eval cache were lost. This repository is the local GPU 3 retrain. The two-run difference (arm −6, Base +2) is single-seed cross-machine noise. This result is a single-seed exploratory outcome, not a confirmatory claim.
Inference notes
- Use this repository's tokenizer and
chat_template.jinja. - NoThink: do not prefill
<think>; if the template acceptsenable_thinking, keep it false. - Stop token ids: 151643 (
<|endoftext|>) and 151645 (<|im_end|>). - Eval sampling used temperature 0.7, top_p 0.8, top_k 20.
Weight checksum
model.safetensors(16,381,517,208 bytes):sha256:fdde336f8a504ce0dae4537b0955bbe660e1072ec093e43cce548e5654958a94
OFFICIAL_MERGE_RECEIPT.json is included for merge provenance. LoRA adapter checkpoints are not in this repo.