Skip to content

DataAttributionEval/DATE-LM-Leaderboard

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

8 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

title DATE-LM Leaderboard
emoji 🏆
colorFrom purple
colorTo blue
sdk gradio
sdk_version 5.23.1
app_file app.py
pinned false

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

DATE-LM Data Attribution Leaderboards

This repo contains the leaderboard code associated with the DATE-LM Paper. The leaderboards are hosted in this HuggingFace Space.

The leaderboards are split into 2 broad categories: Training Data Selection and Applications. Each category contains 3 leaderboards, as indicated below.

  • Pre-Training (10K)
  • Pre-Training (30K)
  • Fine-Tuning
Category Leaderboards
Training Data Selection Pre-Training (10K), Pre-Training (30K), Fine-Tuning
Applications (Toxicity / Bias) Homogeneous, Heterogeneous
Applications (Factual Attribution) Factual Attribution

Details on the tasks corresponding to each leaderboard as well as their code pipelines can be found in the DATE-LM paper and Github repo.

Submission

To submit to the leaderboard: submit via the form in the "Submit Scores" tab on the HuggingFace Space page. This will open up a pull request in this repo. It will need to be merged by a member of the team in order to be displayed in the HuggingFace Space.

Information for Submission include:

  • Influence Scores File
  • Paper/Code/Contact Link
  • Method Name and Category
  • Metrics (dependent on leaderboard chosen)
  • and more

Ranking

Each leaderboard's ranking is based on the values from the metrics, with details specified in the description of each leaderboard. To summarize, the leaderboards are ranked using the following schemes:

Leaderboard Ranking Metric
Pre-Training (10K), Pre-Training (30K) highest score in avg column
Fine-Tuning average of MMLU, GSM8K, and BBH scores
Applications (Toxicity / Bias) highest score in AUPRC column
Applications (Factual Attribution) average of Recall@50 and MRR scores

Repo Files

Overview of Repo files:

  • app file: app.py
    • mappings files: filePaths.py, nameMapping.py
    • Github PR creation file: pr.py
  • submissions storage: submissions folder
    • Note: each submission has its own dedicated folder containing metadata.json and the influence scores
  • leaderboards data: data folder
  • github workflow files:
    • Add submission into leaderboard json upon merge: merge-data.yml, scripts/merge_data.py
    • Sync repo with HuggingFace Space: push-to-hf.yml

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages