modrill/qwen3-4b-think-s1-full-sft
Full-parameter supervised fine-tuning (SFT) of Qwen/Qwen3-4B-Base on the think_s1 curriculum stage (easy + medium code reasoning data). This checkpoint is Stage1 of the think curriculum (checkpoint-1094).
Related models
- Think baseline full SFT: modrill/qwen3-4b-think-baseline-full-sft
- Nothink baseline full SFT: modrill/qwen3-4b-nothink-baseline-full-sft
Training summary
| Field | Value |
|---|---|
| Method | Full SFT (DeepSpeed ZeRO-2) |
| Dataset | think_s1 (easy + medium, 72,555 samples) |
| Chat template | qwen3 |
| Thinking mode | enable_thinking=true |
| Cutoff length | 16384 |
| Packing | true (neat_packing) |
| Epochs | 2 |
| Global batch | 64 (4 GPU × 4 × 4) |
| Learning rate | 1e-5 |
| LR schedule | cosine, warmup 10% |
| Train steps | 1094 |
| Final train loss | ~0.57 |
| Finished | 2026-06-09 |
Eval (EvalScope, release_latest / AIME)
| Benchmark | pass@1 | Config |
|---|---|---|
| LiveCodeBench | 36.06% | t=0.6, p=0.95, max_tokens=16384 |
| AIME24 | 16.67% | same sampling, max_tokens=16384 |
| AIME25 | 3.33% | same sampling, max_tokens=16384 |
Usage
HuggingFace Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "modrill/qwen3-4b-think-s1-full-sft"
tok = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto"
)
vLLM
python -m vllm.entrypoints.openai.api_server \
--model modrill/qwen3-4b-think-s1-full-sft \
--served-model-name think-s1 \
--max-model-len 32768 \
--port 8801
Inference tips
- Use Qwen3 chat template with thinking enabled
- Recommended eval
max_tokens: 16384 (matches training cutoff) - Sampling: temperature=0.6, top_p=0.95, top_k=20
License
Apache 2.0, consistent with the Qwen3 base model license.