Activation_Decoding: Decoding by sharpness inside Large Language Models

Overview

This is the code implementation of the paper: In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation

We discover a pattern associated with hallucinations: correct generations tend to have sharper context activations in the hidden states of the in-context tokens, compared to that of the incorrect generations. Leveraging this signal, we propose an entropy-based metric to quantify the "sharpness " among the in-context hidden states and incorporate it into the decoding process, i.e, use the entropy value to adjust the next token prediction distribution to improve the factuality and overall quality of the generated text.

Setup

conda create --name decode python=3.8
conda activate decode
pip install -e transformers
pip install -r requirements.txt

Experiments

Arguments

Argument	Example	Description
--model-name	daryl149/Llama-2-7b-chat-hf	Specifies the model you want to use, currently we support LLaMA all versions.
--data-path	/path/to/dataset	Path to the dataset file or folder.
--dataset-name	hotpotqa	The dataset for evaluation. Can choose from "natural_questions" or "triviaqa"
--decoding_mode	activation	Choose the decoding mode. Can choose from "baseline", "dola", "activation","activation_dola"
--decoding_strategy	entropy	Add this when the "decoding_mode" is activationor activation_dola. Can choose from "entropy", "single_entropy"
--alpha	0.5	Control the magnitude of activation. For OOD test, we use 0.5/0.8/1.0 for 7B/13B/70B models.
--info_layer	26	The index of layer we use to calculate the entropy. For OOD test, we use 26/34/70 for 7B/13B/70B models.
--output-path	../res/res_hpqa/${model_name}/${decoding_mode}.json	Where to store the output results.
--num-gpus	1	Number of GPUs to use, 1/1/2 for 7B/13B/70B model sizes respectively if you are using 80G GPU cards.
--do-rating		Add this to output evaluation results

In-domain Hyperparameter choice

In the table below, we present the alpha with the info_layer choices for each knowledge-seeking dataset, formatted as alpha/info_layer.

	TriviaQA	HotpotQA	NQ
LLaMa2-7B-chat	1/32	0.8/32	0.2/32
LLaMa2-13B-chat	0.6/32	1/28	0.8/34
LLaMa2-70B-chat	1/70	0.6/66	1/68

TriviaQA, HotpotQA and Natural Questions

cd scripts
bash run_qa.sh # refer to this script for more details

TruthfulQA

Multi Choice

cd scripts
bash run_mctqa.sh # refer to this script for more details

Open generation

To evaluate the open-ended generation result of TruthfulQA, we need to finetune two GPT-3 curie models through OpenAI API:

openai api fine_tunes.create -t finetune_truth.jsonl -m curie --n_epochs 5 --batch_size 21 --learning_rate_multiplier 0.1
openai api fine_tunes.create -t finetune_info.jsonl -m curie --n_epochs 5 --batch_size 21 --learning_rate_multiplier 0.1

After finetuning, we can obtain the finetuned model names by openai api fine_tunes.list | grep fine_tuned_model.

Create a config file gpt3.config.json like this:

{"gpt_info": "curie:ft-xxxxxxxxxx",
"gpt_truth": "curie:ft-xxxxxxxxxx",
"api_key": "xxxxxxx"}

Add the argument --gpt3-config gpt3.config.json for GPT-3 evaluation.

cd scripts
bash run_tqa.sh # refer to this script for more details

Reference Repositories

• Dola: https://github.com/voidism/DoLa

Citation

Please cite our paper if it's helpful to your work!

@inproceedings{
chen2024incontext,
title={In-Context Sharpness as Alerts: An Inner Representation Perspective for Hallucination Mitigation},
author={Chen, Shiqi and Xiong, Miao and Liu, Junteng and Wu, Zhengxuan and Xiao, Teng and Gao, Siyang and He, Junxian},
booktitle={Arxiv},
year={2024},
url={https://arxiv.org/abs/2403.01548}
}