code-think-q8b-20260908

code-think-q8b-20260908

Research checkpoint: Qwen/Qwen3-8B-Base (49e3418fbbbca6ecbdf9608b4d22e5a407081db4) after V4 LoRA SFT on Qwen3-30B-A3B-Thinking-2507 traces, 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/Qwen3-8B-Base (verified from the local base README.md; Apache-2.0 text bundled from the Qwen3-4B-Base LICENSE file on disk).

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 151643 (<|endoftext|>), 151667 (<think>), 151668 (</think>). 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% (3,892,427 / 64,873,788 tokens), weight decay 0.1
• bf16, no packing, no truncation, context 32768
• 2 epochs, physical concat. Endpoint-only score; no checkpoint picking.

Merged weights are the 2-epoch endpoint (step-000921, 64,873,788 assistant tokens). LoRA + B-rows are under adapter/.

Evaluation (DEV256)

256-problem LiveCodeBench-derived dev split. Seed 3407, think mode, no <think> prefill, max generation ~32k, 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.

For context, the same-family 4B base on the current V4 think contract is 63/256 (seed 3407), with a 5-seed band 55.2 ± 4.9 on an earlier bare-base protocol.

Usage

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

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "modrill/code-think-q8b-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,
)
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 plus OFFICIAL_MERGE_RECEIPT.json
• adapter/: LoRA, 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
• MANIFEST.sha256

Optimizer / resume states are not included.