help me, i am doing review summarisation, using a transformer, i have finetuned it…
- User
help me, i am doing review summarisation, using a transformer, i have finetuned it (instruction fine-tuning on 50 cars), using pegasus large, then using in my application via pipeline,
problem is my model is hallucnating, creating details on it's own which were not mentioned, what to do here?
- Assistant
I can help you address the hallucination issue in your review summarization model. Here are several strategies to mitigate hallucination in transformer-based models:
- Data Quality and Preprocessing
- Ensure your training data is high-quality and clean
- Use strict filtering to remove ambiguous or noisy samples
- Increase the diversity of your 50-car training dataset if possible
- Model Configuration Techniques
- Adjust temperature and top-k/top-p sampling parameters
- Lower temperature reduces randomness and hallucination
- Use more conservative decoding strategies
- Implement length penalty to constrain generation
- Training Modifications
- Add a hallucination penalty during fine-tuning
- Use techniques like contrastive learning
- Implement retrieval-augmented generation (RAG)
- Consider using a smaller model like Pegasus Small initially
- Post-processing Strategies
- Implement a fact-checking mechanism
- Cross-reference generated summaries with source text
- Use extractive summarization techniques as a baseline
- Add a confidence threshold for summary acceptance
- Evaluation Metrics
- Use metrics like ROUGE that compare generated vs reference summaries
- Implement human evaluation for hallucination detection
- Track hallucination rate during validation
Would you like me to elaborate on any of these strategies or provide code snippets for implementation?
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.