code

prompts:

• de_gpt4_end2end_prompt_utils.py: prompts used for Section 3 in the paper
• de_prompt_utils.py: prompts for German atomic edit extraction and explanation generation
• zh_prompt_utils.py: prompts for Chinese atomic edit extraction and explanation generation

fine-tune_llama2-7b:

• fine-tune_llama2-7b.sh: parameters for fine-tuning the model
• qlora.py: see source code here

rule_based_screening.py: the heuristic rules for screening out low-level mistakes in atomic edit extraction

SequenceMatcher_rough_edits.py: use SequenceMatcher from difflib to extract rough edits

data

fine-tune_data: the training and test data of LLM fine-tuning for German and Chinese atomic edit extraction. The data is in the format for fine-tuning ChatGPT. Sentence pair is the source and target sentence; list of edits are the rough edits extracted by SequenceMatcher; list of labels are the labels of the edits; content is the gold atomic edits.

human_annotation_data: the anonymized raw human annotation data

Sentence aligner

We modify the paragraph aligner from here to align sentences in the datasets.