/visit_bench_sandbox

Primary LanguagePythonMIT LicenseMIT

Visit Bench Leaderboard

This repo contains code/instructions for updating the visit-bench leaderboard. Link to leaderboard - https://huggingface.co/spaces/mlfoundations/VisIT-Bench-Leaderboard

Commands to add a get ELO Rating for a Single Image model

  1. We need a model predictions csv file with 4 required columns: "instruction", "instruction_category", "image", "model_name", "prediction" where model_name is the model's name. Look at the existing files in the full_model_predictions.
  2. Add this model prediction file to the full_model_predictions folder. Make sure the key used does not overlap with any of the existing keys.
  3. Run make_pairs.py. This will generate all_pairs_with_references_original_set.jsonl (all possible pairs, which we will not be running) and [model_name]_new_pairs_with_references.jsonl (queries that we actually will run in step 5).
  4. Unzip gpt-4_cache.jsonl.zip to get cached gpt-4 judgments using unzip gpt-4_cache.jsonl.zip.
  5. Run
OPENAI_API_KEY=[YOUR OPENAI KEY] python run_all_queries.py --leaderboard_jsonl leaderboard_submission_model_queries/[model_name]_predictions_new_pairs_with_references.jsonl (from step 3)

to add gpt-4 judgments with the new model into gpt-4_cache.jsonl.
6. Run: python evaluate_all_queries.py which will get the judgements gpt-4_head2head.json.
7. Run:

python elo_analysis.py --head2head_file gpt-4_head2head.json

which will output the leaderboard.

Committing to the Leaderboard

Please send your model predictions file and the zipped version of the updated gpt-4_cache.jsonl, from Step 5 above, to the authors at yonatanbitton1@gmail.com and hbansal10n@gmail.com.