ICTNLP/Llama-2-7b-chat-TruthX
ICTNLP/Llama-2-7b-chat-TruthX is a 7 billion parameter Llama-2-7B-Chat model developed by Shaolei Zhang, Tian Yu, and Yang Feng, enhanced with the TruthX method. TruthX is an inference-time technique designed to mitigate hallucinations by editing the model's internal representations in a truthful space. This model is specifically optimized for improving truthfulness, demonstrating an average 20% enhancement on the TruthfulQA benchmark across various LLMs, making it suitable for applications requiring high factual accuracy.
Loading preview...
TruthX: Hallucination Alleviation for LLMs
ICTNLP/Llama-2-7b-chat-TruthX is a 7 billion parameter Llama-2-7B-Chat model that integrates the TruthX method, developed by Shaolei Zhang, Tian Yu, and Yang Feng. TruthX is an innovative inference-time technique designed to enhance the truthfulness of Large Language Models (LLMs) by modifying their internal representations within a "truthful space." This approach effectively mitigates the problem of hallucinations in LLM outputs.
Key Capabilities
- Hallucination Mitigation: Directly addresses and reduces the generation of untruthful information by LLMs.
- Enhanced Truthfulness: Achieves a significant average 20% improvement in truthfulness across 13 advanced LLMs on the TruthfulQA benchmark.
- Seamless Integration: The TruthX method is "baked-in" to this Llama-2-7B-Chat model, allowing it to be used directly like a standard Llama model without additional operational steps.
Good for
- Applications requiring high factual accuracy and reduced hallucination.
- Research and development focused on improving LLM reliability and truthfulness.
- Developers seeking a Llama-2-7B-Chat variant with built-in hallucination alleviation capabilities.
For more technical details, refer to the TruthX paper and the GitHub repository.