qwen3_4b_instruct_2507_sft_gpt5mini_sft_2ep

Fine-tuned Qwen3 4B Instruct on the in-house GPT5-mini supervised fine-tuning (SFT) mixture for two epochs. This checkpoint is exported in safetensors format for safe and fast loading on the Hugging Face Hub.

Model Details

• Architecture: Qwen3ForCausalLM (4B parameters)
• Base model: Qwen/Qwen2.5-4B-Instruct
• Fine-tuning recipe: Supervised fine-tuning (SFT)
• Training run: outputs/qwen3_4b_instruct_2507_sft_gpt5mini_sft_2ep/v0-20251027-193608/checkpoint-90
• Epochs: 2
• Precision: bfloat16 weights stored as safetensors

Intended Use & Limitations

This model is designed for bilingual (Chinese/English) chat-style prompting. Avoid using it for factual or safety-critical tasks without additional alignment and evaluation. Please comply with local regulations and Hugging Face community guidelines when deploying the model.

How to Use

from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "<your-username>/qwen3_4b_instruct_2507_sft_gpt5mini_sft_2ep"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")

prompt = "### Instruction:\n写一段关于量子计算的简介。\n\n### Response:"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=256, do_sample=True, top_p=0.9)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation

No automatic evaluations were run after fine-tuning. Please benchmark on tasks that matter to your use case (e.g., MT-Bench, MMLU, C-Eval) before production deployment.

License

Follows the base model license from Qwen/Qwen2.5-4B-Instruct. Ensure compatibility with your downstream usage.