SeaFill2025/Qwen3-4B-SFT

Qwen3-4B-SFT

Qwen3-4B-SFT is a reasoning-focused model fine-tuned from Qwen3-4B-Base using the verl framework.

In open-source practice, fully reproducible "pre-RL SFT base" releases are rare. This model fills that gap by providing a practical intermediate checkpoint that is Math-forward, Reasoning-focused, and Format-aligned. It is ideal as a clean, warm-start base for Reinforcement Learning (RL) or standalone reasoning tasks.

Configuration Notes

• Model Info: Full-parameter SFT based on Qwen3-4B-Base, optimized for Chain of Thought (COT) reasoning.
• Template: Trained with the Qwen chat template; learns to end responses with <|im_end|> (151645).
• Suggested Configuration:

  {
    "eos_token_id": 151645
  }
  

You may adjust settings according to your training or deployment needs.

Benchmark Snapshot

The following results compare the Base (4B) model against the Qwen3-4B-SFT (this model), highlighting the performance gains in reasoning and mathematics.

Dataset	Base (4B)	Qwen3-4B-SFT (this model)	Improvement (Absolute)
AIME 2024	11.25%	20.8%	+9.55%
AIME 2025	6.46%	19.4%	+12.94%
AMC 2023	31.09%	58.0%	+26.91%
GPQA-Diamond	7.77%	29.1%	+21.33%

• Cluster: MeluXina Supercomputer (LuxProvide)
• Node Config: 4 NVIDIA-A100 GPUs per node.
• Final SFT Run: 12 Node-hours (16× A100 for 3 hours)
• Total R&D Investment: ~700 Node-hours (Includes data ablation, hyperparameter sweeps, and extensive benchmark evaluation.)

Project Links

• Model repository (this page): https://huggingface.co/96kevinli29/Qwen3-4B-SFT
• Dataset card used for SFT: https://huggingface.co/datasets/96kevinli29/SFT-Dataset
• Training code repository: https://github.com/96kevinli29/base-model-sft-verl

Limitations

• Not optimized for factual correctness in all domains
• May still produce hallucinations or unsafe outputs
• Performance is sensitive to prompt style and decoding settings

Citation

If you use this model, please cite this checkpoint, bibTeX for this release :

@misc{qwen3-4b-sft-2026,
  title        = {{Qwen3-4B-SFT}: Supervised Fine-Tuned {Qwen3}-4B for Reasoning},
  author       = {Hongyang Li and Xiao Li and {Sea-Fill Community}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/SeaFill2025/Qwen3-4B-SFT}},
  note         = {Checkpoint trained with verl; warm-start for pre-RL alignment research. Maintained by Sea-Fill Community.}
}

Also cite as appropriate:

• The base model (Qwen3-4B-Base) — use the official Qwen3 / Alibaba citation from its Hugging Face model card.
• The training code repository: https://github.com/96kevinli29/base-model-sft-verl/
• The original source datasets listed in the dataset recipe.