AdaptEval - Framework for evaluation of domain adaptation on a text summarization tasks

Usage

---

1. Using config.json

See config.json for execution configuration

python compile_results.py --config <path_to_json> 

python token_dist_shift_few_shot_zero_shot.py --config <path_to_json> 

python token_dist_shift_model2_model1.py --config <path_to_json> 

python gpt4-eval.py 

Notes

For Reference-based Evaluation

• compile_results.py is intended to compute ROUGE, BERTscore, and domain vocabulary overlap on the provided runs in the input directory through the config.json

For Model Comparisons:

• token_dist_shift_few_shot_zero_shot.py is intended to do Token Distribution Shift analysis for the same model with a zero-shot vs two-shot model.
• token_dist_shift_model2_model1.py is intended to do Token Distribution Shift analysis between two given model. Where Model_2 is the bigger/finetuned model and Model_2 is the smaller

Please provide the parameters in the config file. A sample config file is attached. For running these two scripts, the csv should contain a column prompt_0 that contains the model prompt in the following form You are an expert at summarization. Proceed to summarize the following text. TEXT: {article} SUMMARY: {summary} Proceed to summarize the following text. TEXT: {article} SUMMARY: {summary} … TEXT: {article} SUMMARY:

For Evaluation using GPT-4L

• run_chagpt_eval.ipynb is intended to run summarization eval using the OpenAI API

1.Configuration

Defined in config.py, this script just reads all the configuration from the config.json and passes it on to the resepective evaluation scripts.

2. Metrics

Currently supported metrics:

• BERTscore
• Domain Vocabulary overlap
• ROUGE
• Token Dist Shift
• Evaluation using GPT4

3. Others

• All .csv files in the input directory should have two columns "reference" and "prediction".
• The csv files should follow the name convention {random-name}_MODEL_{model-name}_TASK_{ds}_{task}. This naming convention helps the later group and compare the runs.
• Most importantly, it allows the correct mapping of predictions to their respective domain vocabulary when calculating Domain Vocabulary Overlap.
• During the first run, the domain vocabulary is computed using the domain articles. For later runs, the previously computed domain vocabulary is used.
• Please provide your huggingface token in .env