AttriEval

AttriEval consists of three modules:

• Data: used to load the data and prompts.
• Generation: used to generate answers for a given question and its top-k retrieved documents as context.
• Evalution: used to evaluate the attribution quality and attribution bias.

Installation

Run the following command to install the package:

python setup.py install

AttriEval is tested with Python 3.9.0.

Load Data

The first step is to load one of the benchmarks using load_data_from_disk function.

from attrieval.data_utils import load_data_from_disk, prepare_and_label_data
att_data = load_data_from_disk('./data/nq/')

In the next step, we load proper prompt templates and prepare the data.

Two arguments should be specified for this purpose. rag_mode is set with one of the three rag models (i.e., information settings) are used for answer generation. In Vanilla mode (vanilla), we feed documents to the answer generator LLMs without any labels. In informed (base) and counterfactual mode (cf), relevant and non-relevant documents are labeled with two different set of labels. These labels are set according to the metadata_type which should be set to hai for human versus AI (LLM) authorship. Other types of metadata can be set according to the user preference:

• hai: human versus AI (LLM) authorship.
• gender: male versus female authorship.
• race: black people versus white people authorship.

prep_att_data, sys_prompt_tmplt, user_prompt_tmplt = prepare_and_label_data(att_data, rag_mode='base', metadata_type='hai')

Generation

In order to genereate the answer, three arguments should be specified as follows:

• model_name: the LLM identifier.
• device: the GPU device id for loading the answer generator model and its corresponding tokenizer.
• user_token: if the LLM specified with model_name needs authorized access, hf_token needs to be set with the user's token.

from attrieval.generation import ModelHandler
model_handler = ModelHandler(model_name= "meta-llama/Meta-Llama-3-8B-Instruct",
                             device= "auto",
                             user_token= [user/token])

generation_att_output = model_handler.generate_with_probs(prep_att_data,
                                                 sys_prompt_tmplt,
                                                 user_prompt_tmplt)

Evaluation

Given the ground-truth document which contains the answer, and citations provided in an answer, we use precision and recall to assess the quality of attribution for the provided answer of an LLM. The output of AttriEval generation moduel, i.e., generation_att_output, includes these required elements.

from attrieval import evaluation
evaluation.evaluate_prec_recall(generation_att_output)

CAB (attribution bias) and CAS (attribution sensitivity) can be computed given the generation outputs for different rag modes. For CAB, the output of vanilla mode should be compared against the output of the base mode. For CAS, the output of base mode should be compared against the output of cf mode.

evaluation.cab_evaluation_nq(generation_att_output, generation_att_output_cf)