from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model
model = AutoModelForCausalLM.from_pretrained(
    "Cbgcbg/qwen3-8b-math-full-sft-epoch11-20250725_161659",
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("Cbgcbg/qwen3-8b-math-full-sft-epoch11-20250725_161659")

# Example usage
messages = [
    {"role": "system", "content": "Please reason step by step, and put your final answer within \\boxed{}."},
    {"role": "user", "content": "Find the derivative of f(x) = x^3 + 2x^2 - 5x + 3"}
]

inputs = tokenizer.apply_chat_template(
    messages, 
    tokenize=True, 
    return_tensors="pt", 
    add_generation_prompt=True
)

outputs = model.generate(
    inputs, 
    max_new_tokens=512, 
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(outputs[0][len(inputs[0]):], skip_special_tokens=True)
print(response)

🔗 Related Models

Base Model: unsloth/Qwen3-8B (full precision)
Training Method: Full parameter fine-tuning (92.4% parameters trained)
Progress: Epoch 11/20 (55% complete)
Dataset: Paper&#x27;s Official Dataset (7,110 training samples)
Configuration: Paper&#x27;s exact Stage 1 SFT settings

Paper Source: "A Practical Two-Stage Recipe for Mathematical LLMs"
Training Approach: Full parameter fine-tuning (Stage 1 SFT only)
Final Model: Will be available after 20 epochs complete

📅 Training Timeline

Started: 20250725_161659
Current: Epoch 11/20 checkpoint
Status: Intermediate checkpoint

This model follows the exact configuration from the paper's Stage 1 SFT approach with full parameter fine-tuning for optimal mathematical reasoning performance.