Dataset · xlangai
BRIGHT: Reasoning-Intensive Search Benchmark
xlangai/BRIGHT
An English text retrieval benchmark featuring queries from StackExchange, LeetCode, and math competitions.
What's inside
BRIGHT is described by the authors as the first text retrieval benchmark requiring intensive reasoning for search. Queries are gathered from diverse domains—StackExchange, LeetCode, and math competitions—based on real human data. The language is English, and the size ranges from 1 to 10 million records. Existing retrieval models show low performance on it, with the best nDCG@10 at 22.1.
What you can do with it
- Train or fine-tune a retrieval model for reasoning-based search
- Test a RAG system on complex queries
- Compare models using nDCG@10 on the benchmark
- Tasks
- Text retrieval
- Languages
- English
- Size
- 1M–10M rows
- Downloads this month
- 33.5K
- Likes
- 81
Models trained on it
- mradermacher/Diver-GroupRank-7B-i1-GGUFText ranking 980
- nvidia/llama-nv-embed-reasoning-3bEmbeddings 578
- mradermacher/Diver-GroupRank-7B-GGUFText ranking 438
- mradermacher/llama-nv-embed-reasoning-3b-GGUF 241
- AQ-MedAI/Diver-GroupRank-7BText ranking 58
- AQ-MedAI/Diver-GroupRank-32BText ranking 46
- ForwardAILabs/MRE-T1Embeddings 0
- ForwardAILabs/MQR-A1Text generation 0
How to load it
from datasets import load_dataset
ds = load_dataset("xlangai/BRIGHT")