BTZSC Leaderboard Space

This Space renders the BTZSC leaderboard from result JSON files stored in the dataset repo.

BTZSC evaluates strict zero-shot text classification across:

• 22 English single-label datasets
• 4 task families: sentiment, topic, intent, emotion
• 4 major model families: embedding models, NLI cross-encoders, rerankers, and instruction-tuned LLMs

Ranking policy:

• Primary ranking metric: macro-F1
• Additional displayed metrics: accuracy, macro-precision, macro-recall

Data source configuration:

• Default dataset repo: btzsc/btzsc-results
• Override with env var: BTZSC_RESULTS_REPO

Useful links:

• Paper (OpenReview): https://openreview.net/forum?id=IxMryAz2p3
• Eval harness (GitHub): https://github.com/IliasAarab/btzsc
• Results dataset: https://huggingface.co/datasets/btzsc/btzsc-results

How to Add a New Leaderboard Entry

This Space is read-only: it displays whatever is present in the results dataset.

To add a new entry, submit a valid JSON result file to:

• https://huggingface.co/datasets/btzsc/btzsc-results
• Required path: results/<model_type>/<model-name>.json

Submission options:

• Web UI upload on the dataset page
• Git workflow (direct push or fork + PR)
• huggingface_hub API upload with create_pr=True (PR-based)

Full submission requirements are documented in:

• https://huggingface.co/datasets/btzsc/btzsc-results/blob/main/SUBMISSION.md

Once merged, the new entry appears in this Space on refresh/update.

For local testing, set:

• BTZSC_RESULTS_LOCAL_DIR=/path/to/results_repo

Then run:

python app.py