Fine-grained-Attribution

Source code for the ACL 2024 Findings paper "Learning Fine-Grained Grounded Citations for Attributed Large Language Models"

Requirements

The required Python packages are listed in requirements.txt. You can create a new conda environment, then run the following command to install them.

conda create -n front python=3.10
conda activate front
pip install -r requirements.txt

Data

You can directly download both the raw and processed dataset from this Google Drive link.

Training

We use 4xA100 80G GPUs for the two-stage training.

Stage1: Grounding Guided Generation

cd training/stage1_grounding_guided_generation
sh train_sft.sh

Stage2: Consistency-Aware Alignment

cd training/stage2_consistency_aware_alignment
sh train_dpo.sh

Evaluation

For evaluation, please refer to ALCE.

Bugs or Questions?

If you have any questions related to the code or the paper, feel free to email [email protected]