DiverseAgentEntropy

This is the repo to reproduce the paper DiverseAgentEntropy, a novel method for quantifying the uncertainty in the factual parametric knowledge of Large Language Models (LLMs).

Setup

Please install the packages in requirements.txt

Data preparation

You should prepare QA datasets in the following format under ../data:

[
  {
    "question": "<string: the input question>",
    "gold_answer": "<string: the authoritative or ground-truth answer>",
  },
  {
    "question": "<string: the input question>",
    "gold_answer": "<string: the authoritative or ground-truth answer>",
  },
  // ... more entries ...
]

API inference

We are currently using the AWS Bedrock platform, but you can switch to any other API call method by updating the ../code/utils.py file accordingly.

Diverse agent question generation

We first generate varied questions about the same underlying original query with different contexts.

python -m code.question_generation.pipeline_question_generation --dataset_name=dataset_name --model_name=model_name 

We then select n questions as the final questions for agent interaction.

python -m code.question_generation.pipeline_question_selection --dataset_name=dataset_name --model_name=model_name 

Agent Interaction

We use the selected diverse questions to encourage agent interactions to further reveal the model's knowledge of the original query.

python -m code.agent_interaction.pipeline_agent_interaction  --dataset_name=dataset_name --model_name=model_name --mode=origin

We then implement the abstention policy to get the uncertainty score for each original query.

python -m code.evaluation.agent_evaluation  --dataset_name=dataset_name --model_name=model_name --mode=origin

Baseline

We use the following code to run the self-consistency-based SemanticEntropy baseline.

python -m code.baseline.vanilla_qa  --dataset_name=dataset_name --model_name=model_name

We evaluate the baseline:

python -m code.evaluation.vanilla_evaluation  --dataset_name=dataset_name --model_name=model_name

Evaluation

Finally, we evaluate different methods and draw the AR-curve.

python -m code.evaluation.draw_figure  --dataset_name=dataset_name --model_name=model_name

License

This project is licensed under the Apache License 2.0 - see the LICENSE file for details.

🐛 Issues

If you encounter any issues or have questions, please open an issue on GitHub or contact the authors.