DavidOKBm/MathThink-Qwen-3.5-4B

Qwen3.5-4B Math Fine-Tuned (Nemotron-SFT-Math-v3)

This model is a fine-tuned version of Qwen3.5-4B, explicitly optimized for complex mathematical reasoning and Chain-of-Thought (CoT) problem solving. It was fine-tuned using the Nemotron-Math-v3 dataset with Parameter-Efficient Fine-Tuning (PEFT/LoRA).

Model Details

• Base Model: Qwen/Qwen3.5-4B
• Fine-Tuning Dataset: nvidia/Nemotron-SFT-Math-v3
• Methodology: LoRA (Rank = 64, Alpha = 32 or Alpha = 16). The lora_alpha scaling is specifically tuned to prevent catastrophic forgetting, ensuring the model retains conversational abilities while significantly enhancing mathematical logic.
• Quantization: Safetensor format (F16) and GGUF formats (Q8_0)

Recommended Generation Parameters

Because this model leverages extensive Chain-of-Thought reasoning to solve math problems, the following generation parameters are highly recommended for the best performance:

{
  "temperature": 1.0,
  "top_p": 0.95,
  "repetition_penalty": 1.1
}

Note: A repetition_penalty of 1.1 is crucial to prevent the base model from occasionally falling into infinite generation loops on extremely long context windows.

Use Cases

• Resolving complex math word problems (GSM8K).
• Higher-level mathematical reasoning (MATH, AIME).
• Step-by-step logic tracking and proofs.