ai-forever/RUMBA-leaderboard-data

RUMBA Leaderboard Data

This dataset repository stores leaderboard data and submission metadata for the RUMBA benchmark.

Files

• leaderboard.csv — Accepted submissions shown in the public leaderboard
• submissions/ — Uploaded submission artifacts
  • <submission_id>/original_submission.zip — Original ZIP archive uploaded by the participant
  • <submission_id>/results.json — Extracted LightEval results file
  • <submission_id>/category_avg_score_<ru/en>.json — Extracted RUMBA category scores file
• pending/ — Submissions waiting for review
  • <submission_id>.json — Pending submission metadata
• accepted/ — Archived accepted submissions
  • <submission_id>.json — Accepted submission metadata
• rejected/ — Archived rejected submissions
  • <submission_id>.json — Rejected submission metadata

Leaderboard Columns

• rank — Position in leaderboard (per language)
• method — Method name
• hf_username — Hugging Face username (via OAuth)
• language — Dataset language (ru/en)
• embedder_model — Embedding model used
• answer_model — Answer generation model used
• details — Short public details about the solution and key configuration choices
• overall_score — Overall RUMBA Score (llm-judge-accuracy)
• single_score — Score on single-session questions (weighted llm-judge-accuracy)
• multi_score — Score on multi-session questions (weighted llm-judge-accuracy)
• temporal_score — Score on temporal questions (weighted llm-judge-accuracy)
• atemporal_score — Score on atemporal questions (weighted llm-judge-accuracy)
• extraction_score — Score on Extraction questions (weighted llm-judge-accuracy)
• reasoning_score — Score on Reasoning questions (weighted llm-judge-accuracy)
• abstention_score — Score on Abstention questions (weighted llm-judge-accuracy)
• setting — Evaluation setting
• model_url — Link to model
• code_url — Link to code repository
• paper_url — Link to paper/report
• date_utc — Submission date (UTC)
• submission_id — Unique submission ID
• status — Status (accepted/pending/rejected)

Submission Metadata

Each submission metadata JSON stores additional information used for review and audit, including the submitted ZIP archive path, extracted artifact paths, validation diagnostics, hashes, links provided by the participant, and the free-form method details.

The public leaderboard is generated from accepted submissions only. Pending and rejected submissions are kept for review history and auditability.