metadata

license: gpl-3.0

TruthX: Alleviating Hallucinations by Editing Large Language Models in Truthful Space

Shaolei Zhang, Tian Yu, Yang Feng*

TruthX is an inference-time method to elicit the truthfulness of LLMs by editing their internal representations in truthful space, thereby mitigating the hallucinations of LLMs. On the TruthfulQA benchmark, TruthX yields an average enhancement of 20% in truthfulness across 13 advanced LLMs.

TruthfulQA MC1 accuracy of TruthX across 13 advanced LLMs

This repo provide Llama-2-7B-Chat-TruthX, a Llama-2-7B-Chat model with baked-in TruthX model. You can directly download this baked-in model and use it like standard Llama, no additional operations are required.

Quick Starts

Inference with Llama-2-7B-Chat-TruthX:

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

llama2chat_with_truthx = "ICTNLP/Llama-2-7b-chat-TruthX"
tokenizer = AutoTokenizer.from_pretrained(llama2chat_with_truthx, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(llama2chat_with_truthx, trust_remote_code=True,torch_dtype=torch.float16).cuda()

question = "What are the benefits of eating an apple a day?"
encoded_inputs = tokenizer(question, return_tensors="pt")["input_ids"]
outputs = model.generate(encoded_inputs.cuda())[0, encoded_inputs.shape[-1] :]
outputs_text = tokenizer.decode(outputs, skip_special_tokens=True).strip()
print(outputs_text)

Please refer to GitHub repo and our paper for more details.