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

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Apr 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

varunchundru/dpo-qwen2.5-0.5b-halueval is a 0.5 billion parameter Qwen2.5-Instruct model fine-tuned by Varun Chundru using Direct Preference Optimization (DPO). It is specifically designed to reduce hallucination across grounded QA, summarization, and dialogue tasks, achieving a 55.9% relative reduction in overall hallucination rate. This model excels in grounded question answering and summarization by generating more faithful responses based on provided context.

Loading preview...

Overview

This model, dpo-qwen2.5-0.5b-halueval, is a 0.5 billion parameter Qwen2.5-Instruct model fine-tuned by Varun Chundru using Direct Preference Optimization (DPO). It was developed as part of a project on domain-specific hallucination detection in large language models. The primary goal of this model is to mitigate hallucination in generated text, particularly for grounded tasks.

Key Capabilities & Performance

  • Significant Hallucination Reduction: Achieves a 55.9% relative reduction in overall hallucination rate compared to its base model, lowering it from 85.5% to 37.7% on the HaluEval benchmark.
  • Strong Performance in QA and Summarization: Demonstrates a 79.7% relative reduction in hallucination for QA and 100% elimination for summarization on the test set.
  • DPO Fine-tuning: Trained on 21,000 preference pairs derived from the HaluEval benchmark's QA, dialogue, and summarization subtasks.
  • Mitigation Component: Designed to work as a mitigation component within a larger hallucination detection and mitigation pipeline.

Intended Use Cases

  • Grounded Question Answering: Ideal for scenarios requiring answers strictly based on provided knowledge.
  • Document Summarization: Suitable for generating summaries that accurately reflect the source document without inventing facts.
  • Research: Valuable for studies on hallucination mitigation through preference learning.
  • Paired with Detector: Can be used in conjunction with varunchundru/hallucination-detector-deberta for a complete hallucination management system.

Limitations

  • Weak Dialogue Performance: Shows minimal improvement in dialogue tasks, with a 94.2% hallucination rate post-DPO, indicating challenges with multi-turn conversations at this scale.
  • Training Scope: Trained for only one epoch on a 0.5B parameter model; further training or a larger base model could yield better results.

Finetunes
1 models