SUBMISSION.md

Submitting to the BTZSC Leaderboard

Requirements

• Your model must be publicly available on the Hugging Face Hub.
• Results must be generated using the official btzsc harness.
• All 22 datasets must be evaluated (partial submissions are not accepted).

Steps

1) Install the harness

pip install btzsc

2) Run the evaluation

btzsc evaluate \
  --model <your-hf-model-id> \
  --type <embedding|nli|reranker|llm> \
  --output-json results/<model_type>/<model-name>.json

Example:

btzsc evaluate \
  --model intfloat/e5-base-v2 \
  --type embedding \
  --output-json results/embedding/e5-base-v2.json

3) Validate locally

python validate.py results/<model_type>/<model-name>.json

4) Open a Pull Request

All submission actions in this step target the results dataset repo btzsc/btzsc-results: https://huggingface.co/datasets/btzsc/btzsc-results.

Required destination path:

• results/<model_type>/<model-name>.json

Example:

• results/embedding/e5-base-v2.json

Choose one submission workflow:

1. Web UI (no clone required)

• Open https://huggingface.co/datasets/btzsc/btzsc-results and upload the JSON in Files and versions at the required path.
• If you do not have write access, fork btzsc/btzsc-results and open a PR to btzsc/btzsc-results.

2. Git workflow (clone/fork + push)

• Clone or fork https://huggingface.co/datasets/btzsc/btzsc-results.
• Add your JSON at the required path.
• Push directly (if you have write access) or push to your fork and open a PR to btzsc/btzsc-results.

3. API workflow (huggingface_hub, PR-based)

• Authenticate first (huggingface-cli login or HF_TOKEN).
• Use create_pr=True against repo_id btzsc/btzsc-results to open a PR branch programmatically.
• If PR creation is restricted for your account, upload to your fork and open a PR to btzsc/btzsc-results manually.

For every PR, include:

• The result JSON file.
• A short model description (architecture/training notes).
• Confirmation that model weights are public on Hugging Face Hub.

Merge checks

• JSON passes schema validation (CI).
• Model is public on Hugging Face Hub.
• Scores are plausible (spot checks on difficult datasets).
• evaluation.btzsc_version matches a released version.

Data contamination disclosure

If any BTZSC dataset was explicitly used in your model's training data, disclose it in your PR description. We track this for transparency.