code-think-q4b-20260908

code-think-q4b-20260908

Research checkpoint: Qwen/Qwen3-4B-Base (906bfd4b4dc7f14ee4320094d8b41684abff8539) after V4 LoRA SFT on a Mix Distillation payload, then merged to full weights.

This is a research checkpoint, not a product. Single-seed diagnostic numbers only. Do not treat DEV256 as a leaderboard claim.

License: Apache-2.0, inherited from Qwen/Qwen3-4B-Base (verified from the local base README.md / LICENSE).

Base, teacher, and data

Recipe

• LoRA r64 / α128, dropout 0.0, seven projections: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Embeddings / lm_head frozen except two-sided trainable B-rows for special tokens 151643 (<|endoftext|>), 151667 (<think>), 151668 (</think>). Qwen3-4B-Base embeddings were tied; training untied them. Token ids from adapter/TOKEN_ROWS_META.json.
• Assistant supervised tail: <|endoftext|> (151643)
• LR 1e-4, AdamW (β 0.9/0.95), cosine over assistant-token dose, warmup 6% (3,890,907 / 64,848,450 tokens), weight decay 0.1 (LoRA; row WD 0)
• bf16 activations, fp32 trainable masters, no packing, no truncation, context 32768
• 2 epochs, physical concat of the 1-epoch mix (same order). Endpoint-only score; no checkpoint picking.

Merged weights in this repo are the 2-epoch endpoint (step-000920, 64,848,450 assistant tokens). The LoRA adapter and B-row file are under adapter/.

Evaluation (DEV256)

256-problem LiveCodeBench-derived dev split. Seed 3407, think mode, no <think> prefill, max generation ~32k (model context 32768), sandbox-verified pass@1. Temperature 0.6, top-p 0.95, top-k 20.

Cap = generations that hit the 32k length limit without closing </think>.

Single seed. These are research checkpoints, not product scores.

Why Mix (sibling V4 controls)

Pure strong-teacher traces hurt this 4B student; a same-size teacher and the 0.2:0.8 mix did not:

That pattern matches Li et al. 2025 (arXiv:2502.12143): small students often learn worse from much stronger reasoners than from a mixed or same-size teacher.

Usage

Merged full weights; no PEFT required at inference. Think mode: pass enable_thinking=True and do not prefill <think>.

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "modrill/code-think-q4b-20260908"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype="bfloat16", device_map="auto"
)

system = (
    "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."
)
messages = [
    {"role": "system", "content": system},
    {"role": "user", "content": problem_statement},
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,  # think mode; no <think> prefill
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
output = model.generate(
    **inputs,
    max_new_tokens=32768,
    do_sample=True,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=False))

Stop ids used in the official eval: 151643 (<|endoftext|>), 151645 (<|im_end|>).

Repo layout

• Root: merged HF weights (config.json, model.safetensors, tokenizer, generation_config.json, chat_template.jinja) plus the merge record OFFICIAL_MERGE_RECEIPT.json
• adapter/: LoRA (adapter_config.json, adapter_model.safetensors), two-sided B-rows (token_rows_both_sides.safetensors), TOKEN_ROWS_META.json, checkpoint MANIFEST.json
• provenance/: train RUN_IDENTITY.json, TRAINING_CONFIG.json, POLICY.json; DEV256 COMPLETE.json; mix builder and MQ0 near-dup report
• MANIFEST.sha256: sha256 of every uploaded file

Optimizer / resume states are not included.