Sentence Selection ORPO LoRA

Training Details

• Base Model: Qwen3-30B-A3B (via SFT fine-tuned version)
• Method: ORPO (Odds Ratio Preference Optimization)
• Task: Select relevant sentence IDs from academic texts to support claims
• F1 Score: 0.247 on holdout set

Usage

Key Findings

1. Format-consistent DPO pairs essential for ORPO training
2. 2 epochs optimal (more causes overfitting)
3. Noise augmentation fixes positional bias in training data