code-nothink-q4b-20260908

code-nothink-q4b-20260908

Formerly Qwen3-4B-Base-code-v4-nothink (HF redirect).

Arm ID: Q4B-NOTHINK
Run ID: t30b2507-q4b-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).

Single-seed exploratory result, not a preregistered confirmatory claim.

Base model

• Hugging Face: Qwen/Qwen3-4B-Base
• Revision: 906bfd4b4dc7f14ee4320094d8b41684abff8539

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: mathbox; eval tokenizer renderer qwen3-v4-tail151643

Evaluation

• Suite: official LiveCodeBench DEV256
• Seed 3407, mode NoThink, max_model_len 8192, vLLM 0.28.0
• Metric: sandbox pass@1 = passed / 256
• This arm: 67/256 (26.2%), caps 99
• Base (Qwen/Qwen3-4B-Base, same NoThink protocol): 36/256 (14.1%), caps 48
• McNemar exact p = 7.8e-07

Inference notes

• Use this repository's tokenizer and chat_template.jinja.
• NoThink: do not prefill <think>; if the template accepts enable_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 (8,822,894,520 bytes): sha256:4309b112b8635695282a1c18f7d9e301862ef028c83e2b000a34caa08b47ef1f

OFFICIAL_MERGE_RECEIPT.json is included for merge provenance. LoRA adapter checkpoints are not in this repo.