Reduce hallucinations with Bespoke-Minicheck

2024-09-17 · Source: Ollama Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Intermediate, quick

Summary

Bespoke-Minicheck, a new grounded factuality checking model developed by Bespoke Labs, is now available in Ollama as of September 18, 2024. This model is designed to detect and reduce hallucinations in responses generated by other language models. It operates by taking a "Document" (factual information) and a "Claim" (generated output), then verifies if the document supports the claim, outputting "Yes" or "No". Bespoke-Minicheck is particularly effective in Retrieval Augmented Generation (RAG) applications, where it can serve as a post-processing step to ensure LLM responses are grounded in the provided context. Users can download and run the model via `ollama run bespoke-minicheck` and provide document-claim pairs for verification.

Key takeaway

For AI Engineers building RAG applications, integrating Bespoke-Minicheck as a post-processing step can significantly enhance the factual accuracy of your LLM outputs. You should consider adding this model to your existing pipelines to automatically verify generated claims against retrieved context, thereby reducing hallucinations and improving overall system reliability. Explore the provided GitHub examples for quick implementation.

Key insights

Bespoke-Minicheck verifies claims against documents to reduce LLM hallucinations, especially in RAG.

Principles

• Ground LLM outputs in provided context.
• Fact-check claims against source documents.

Method

Bespoke-Minicheck takes a factual document and a generated claim, then outputs "Yes" if the document supports the claim or "No" if it does not, enabling post-processing hallucination detection.

In practice

• Integrate into RAG pipelines for post-processing.
• Use `ollama run bespoke-minicheck` to start.
• Provide "Document" and "Claim" for verification.

Topics

• Grounded Factuality
• Hallucination Reduction
• Retrieval-Augmented Generation
• LLM Fact-Checking
• Ollama

Code references

• ollama/ollama
• Liyan06/MiniCheck

Best for: AI Engineer, Machine Learning Engineer, Data Scientist

Related on AIssential

• Hallucinations in LLMs: A Deep Technical Dive into Causes, Detection, and Mitigation — LLM hallucinations are inherent to their probabilistic nature, driven by training objectives that prioritize fluency over factual accuracy.
• MARCH: Multi-Agent Reinforced Self-Check for LLM Hallucination — MARCH uses multi-agent reinforcement learning and information asymmetry to reduce LLM hallucination in RAG systems.
• RAG Doesn’t Kill Hallucinations. — RAG transforms hallucinations into confidently cited errors; robust architectural defenses are essential.
• 5 Practical Techniques to Detect and Mitigate LLM Hallucinations Beyond Prompt Engineering — LLM hallucinations are a system problem requiring multi-layered detection and mitigation beyond prompt engineering.
• [D] Self-Promotion Thread — The AI/ML community actively develops diverse tools, from cost-efficient LLM interaction to advanced model evaluation and production monitoring.
• Reducing Hallucination in Vision-Language Models via Stage-wise Preference Optimization under Distribution Shift — Stage-wise preference optimization with targeted data construction effectively reduces VLM hallucination by focusing on difficult grounding boundaries.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Ollama Blog.