Encountered difficulties in reproducing the results on TruthfulQA

[zlkqlyc]

Hi, thanks for sharing your work! I’m having some trouble reproducing the results and would appreciate your help. For the TruthfulQA Multiple-Choice task, I got these results when evaluating LLMs + TruthX, MC1: 0.5018, MC2: 0.7050, MC3: 0.3905. These are different from the results in the paper (54.22%, 73.90%, 44.37%). How can I adjust to match the results from the paper? Thanks!

For your reference, I’ve attached the relevant scripts and code snippets: We used the model file from https://huggingface.co/ICTNLP/Llama-2-7b-chat-TruthX which does not implement two-fold validation. As a result, we made a few minor adjustments to the scripts and llm.py file, as detailed below:

1. Script content

• Based on scripts/truthfulqa.mc.truthx.sh:

export CUDA_VISIBLE_DEVICES=0

ROOT=path_to_truthx_dir EXP_ROOT=$ROOT/results model_path=path_to_llm # e.g. Llama-2-7b-chat-hf

truthx_model1=truthx_models/Llama-2-7b-chat-hf/truthx_model.fold1.pt truthx_model2=truthx_models/Llama-2-7b-chat-hf/truthx_model.fold2.pt

strength=4.5 layers=10

python3 $ROOT/scripts/truthfulqa_mc_truthx.py
--model-path $model_path
--truthx-model $truthx_model1
--truthx-model2 $truthx_model2
--two-fold True
--data-yaml data/truthfulqa_data_fold1.yaml
--edit-strength $strength --top-layers $layers
--fewshot-prompting True
--output-dir $EXP_ROOT/truthfulqa_mc_truthx/llama-2-7b-chat.truthx

• Our Modifications:

export CUDA_VISIBLE_DEVICES=6

ROOT=. EXP_ROOT=$ROOT/results model_path="/app/model_download/Llama-2-7b-chat-hf" truthx_model1=/app/baseline/TruthX/truthx_models/Llama-2-7b-chat-hf/truthx_model.pt

strength=4.5 layers=10

python3 $ROOT/scripts/truthfulqa_mc_truthx.py
--model-path /app/model_download/Llama-2-7b-chat-hf
--truthx-model $truthx_model1
--edit-strength $strength --top-layers $layers
--fewshot-prompting True
--output-dir $EXP_ROOT/truthfulqa_mc_truthx/llama-2-7b-chat.truthx

2. In the llm.py file, we added the get_lprobs_with_ae function.

• Based on get_lprobs_with_ae_2fold function:

outputs, past_key_values, hidden_states = self.model( input_ids, output_hidden_states=True, truthx_model=( self.truthx if idx not in self.fold1_data else self.truthx2 ), ).values()

• Our Modifications:

outputs, past_key_values, hidden_states = self.model( input_ids, output_hidden_states=True, truthx_model=self.truthx, ).values()

• The rest of the code is the same as get_lprobs_with_ae_2fold function.