varunchundru/dpo-qwen2.5-0.5b-halueval

Overview

This model, dpo-qwen2.5-0.5b-halueval, is a 0.5 billion parameter Qwen2.5-Instruct model fine-tuned with Direct Preference Optimization (DPO). Its primary goal is to mitigate hallucination across various NLP tasks, including grounded Question Answering (QA), dialogue, and summarization. The training utilized preference pairs derived from the HaluEval benchmark, contrasting correct responses against hallucinated ones.

Key Capabilities & Performance

• Significant Hallucination Reduction: Achieves a 55.9% relative reduction in overall hallucination rate compared to its base model.
• Exceptional QA & Summarization Faithfulness: Demonstrates a 79.7% relative reduction in hallucination for QA and effectively eliminates hallucination (100% relative reduction) in summarization tasks on the test set.
• Mitigation Component: Designed as a mitigation component within a larger hallucination detection and mitigation pipeline, intended to work alongside a separately fine-tuned DeBERTa detector.

Intended Use Cases

• Grounded QA and Summarization: Ideal for applications where factual accuracy and faithfulness to provided context are critical.
• Research on Hallucination Mitigation: Useful for studies exploring preference learning techniques to combat LLM hallucination.
• Paired with Detectors: Functions as a generation-time mitigation component when integrated with hallucination detection systems like varunchundru/hallucination-detector-deberta.

Limitations

• Weak Dialogue Performance: Shows minimal improvement in reducing hallucination for multi-turn dialogue, suggesting this task may require more extensive training or a larger model.
• Scale and Training: Trained for only one epoch on a 0.5B parameter model, indicating potential for further improvement with more resources.