Qwen3-0.6B-OpenR1-SFT

Qwen3-0.6B-OpenR1-SFT

Qwen3-0.6B-OpenR1-SFT is a Qwen/Qwen3-0.6B-Base model fine-tuned with the Open-R1 SFT distillation recipe on the open-r1/Mixture-of-Thoughts dataset.

Base Model

• Base model: Qwen/Qwen3-0.6B-Base
• Architecture: Qwen3ForCausalLM
• Max position embeddings: 32768
• Intended use: reasoning-oriented text generation

Training Recipe

This run used the Open-R1 supervised distillation recipe derived from recipes/OpenR1-Distill-7B/sft/config_distill.yaml with the base model overridden to Qwen/Qwen3-0.6B-Base and the Hub target set to RoyceLu/Qwen3-0.6B-OpenR1-SFT.

Key settings:

• Dataset: open-r1/Mixture-of-Thoughts (all config)
• Epochs: 5
• Max sequence length: 32768
• Learning rate: 4e-5
• Per-device train batch size: 2
• Gradient accumulation steps: 8
• Precision: bf16
• Attention implementation: flash_attention_2
• Liger kernel: enabled

Data

• Training dataset: open-r1/Mixture-of-Thoughts
• Training samples seen: 349,317

Hardware And Runtime

• Hardware: 1 x node, 8 x NVIDIA H100 80GB
• Training runtime: 168,141.7s (1d 22h 42m 22s)
• Train steps: 13,650
• Final train loss: 0.6757

Evaluation

Benchmarks were evaluated with lighteval + vLLM on single H100 GPUs. For all runs below, both baseline and fine-tuned models used the same inference settings:

• max_model_length=8192
• temperature=0.6
• top_p=0.95
• use_chat_template=true

max_new_tokens=2048

max_new_tokens=4096

Usage

from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="RoyceLu/Qwen3-0.6B-OpenR1-SFT",
    device="cuda",
)

messages = [{"role": "user", "content": "Solve x^2 - 5x + 6 = 0."}]
output = pipe(messages, max_new_tokens=512, return_full_text=False)
print(output[0]["generated_text"])

Citation

If you use this model in research, evaluations, or derivative model releases, please cite this model together with Qwen3 and Open-R1.

@misc{lu2026qwen3openr1sft,
  author       = {Royce Lu},
  title        = {Qwen3-0.6B-OpenR1-SFT},
  year         = {2026},
  howpublished = {Hugging Face model repository},
  note         = {Fine-tuned from Qwen/Qwen3-0.6B-Base using the Open-R1 supervised distillation recipe}
}