exp007-dpo-fixed

DPO with length bias fixed dataset (sample exclusion method).

Training Configuration

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Method: DPO (Direct Preference Optimization)
• Dataset Fix: Excluded samples where Rejected CoT > 1.5x Chosen CoT
• Epochs: 1
• Learning rate: 1e-07
• Beta: 0.1
• Max sequence length: 1024