Add dataset: Bright-Pro

[Samoed]

dataset link on Hugging dataset

https://huggingface.co/datasets/yale-nlp/Bright-Pro

Arxiv link

https://arxiv.org/abs/2605.04018

Description of the dataset

Congratulations with release @yilunzhao! Would be great to integrate your benchmark into mteb

[yilunzhao]

Hi @Samoed, thanks for opening the issue! I'm one of the BRIGHT-Pro authors and have prepared a PR to integrate the benchmark: #4651.

What's added
Seven per-domain retrieval tasks (one per StackExchange domain in the paper) under mteb/tasks/retrieval/eng/bright_pro_retrieval.py:
BrightProBiologyRetrieval, BrightProEarthScienceRetrieval, BrightProEconomicsRetrieval, BrightProPsychologyRetrieval, BrightProRoboticsRetrieval, BrightProStackoverflowRetrieval, BrightProSustainableLivingRetrieval.

Design choices

• Mirrors the per-domain pattern from bright_v1_1_retrieval.py (each domain is its own AbsTaskRetrieval subclass, main_score=ndcg_at_10, BRIGHT-style per-domain prompt for comparability).
• Uses the documents and examples HF configs from yale-nlp/Bright-Pro and exposes binary qrels from gold_ids.
• BRIGHT-Pro also ships an aspects config with reasoning-aspect annotations and per-aspect Likert weights. Since standard MTEB retrieval consumes binary qrels only, this PR keeps the eval path standard and leaves aspects available on the Hub for users who want aspect-aware metrics (weighted recall, α-nDCG).
• examples["id"] is int64 on the Hub; coerced to str to match corpus doc-ids and MTEB's qid conventions.
• License set to mit (per the dataset card); pinned dataset revision included.

Validation

• ruff check + ruff format --check clean.
• All 7 tasks resolve via mteb.get_tasks(...); biology subset loads with 59,513 corpus docs / 103 queries / 103 qrels and sampled gold doc-ids resolve in the corpus.
• BibTeX passes the citation-formatting test.

Could you please review #4651? Happy to address any feedback. Thanks!