Shypd.aiThe largest open index of AI toolshonest_llama is an AI Agents & Automation tool that enhances the truthfulness of large language models through Inference-Time Intervention (ITI). It provides code and models for applying ITI to LLaMA, Alpaca, and Vicuna models.
Trending
About
honest_llama provides the code and resources for implementing Inference-Time Intervention (ITI), a technique designed to improve the truthfulness of large language models (LLMs). This repository demonstrates how to apply ITI and various baseline methods to models like LLaMA, Alpaca, and Vicuna. ITI works by shifting model activations during inference across specific attention heads, significantly boosting truthfulness on benchmarks like TruthfulQA. The tool offers pre-trained ITI baked-in models on HuggingFace and supports activation editing. It is computationally inexpensive and data-efficient, requiring only a few hundred examples to locate truthful directions, unlike methods that need extensive annotations. The project also highlights a tradeoff between truthfulness and helpfulness, allowing users to balance these aspects by tuning intervention strength. It includes evaluation workflows and integration with the pyvene library for streamlined intervention sharing.
Capabilities
Pricing & Plans
Open Source
Free
FAQs
Trending
