Bright-Pro

Evaluation code for the Bright-Pro benchmark (ACL 2026): an expert-annotated extension of BRIGHT for reasoning-intensive retrieval, with both static and agentic-search evaluation protocols.

• 📊 Dataset: yale-nlp/Bright-Pro
• 🤖 Retriever: yale-nlp/RTriever-4B
• 📁 Collection: Bright-Pro on HuggingFace

What's in this repo

.
├── bright_pro_data.py         # tiny loader; defaults to load_dataset("yale-nlp/Bright-Pro")
├── retrieval/                 # static retrieval (α-nDCG@k, A-Recall@k, NDCG@k, Recall@k)
│   ├── run.py                 # main entry — embed corpus, retrieve, score
│   ├── retrievers.py          # all retrievers reported in the paper
│   ├── metrics.py             # α-nDCG / A-Recall / NDCG / Recall implementations
│   ├── qwen3_embedding.py     # sentence-transformers wrapper for the Qwen3-Embedding family
│   └── configs/<retriever>/<task>.json   # per-(retriever, task) instructions
├── agentic_retrieval/         # agentic search loop + LLM-as-Judge
│   ├── search_agent/          # OpenAI / Qwen agents, fixed-round and adaptive-round
│   ├── searcher/              # in-loop search tool (one class per retriever family)
│   ├── scripts_evaluation/    # judge.py, α-nDCG aggregator, AER, ...
│   └── configs/<retriever>/<task>.json
└── agentic_sample_ids.json    # fixed 175-query subset (25 / task, seed=42) for agentic eval

Retrievers reported in the paper

Category	Retriever	HF id
Lexical	BM25	—
General-purpose	GritLM-7B	GritLM/GritLM-7B
	Instructor-XL (1.5B)	hkunlp/instructor-xl
	GTE-Qwen2-7B-Instruct	Alibaba-NLP/gte-Qwen2-7B-instruct
	Qwen3-Embedding-8B	Qwen/Qwen3-Embedding-8B
	EmbeddingGemma-300M (static-only)	google/embeddinggemma-300m
	OpenAI text-embedding-3-Large (static-only)	OpenAI API
Reasoning-intensive	ReasonIR-8B	reasonir/ReasonIR-8B
	DIVER-Retriever-4B	AQ-MedAI/Diver-Retriever-4B
	DIVER-Retriever-4B-1020	AQ-MedAI/Diver-Retriever-4B-1020
	BGE-Reasoner-Embed-Qwen3-8B	BAAI/bge-reasoner-embed-qwen3-8b-0923
	INF-Retriever-v1-Pro (static-only)	infly/inf-retriever-v1-pro
	RTriever-4B (ours)	yale-nlp/RTriever-4B

The three retrievers tagged "static-only" are evaluated only in the static setting (paper Table 1); the agentic protocol (Tables 2, 3) uses the remaining 10.

Agents and judge

Role	Model
Search agent (primary)	gpt-5-mini-08-07
Search agent (alternative)	Qwen3.5-122B-A10B
LLM-as-Judge	gpt-5

Setup

conda env create -f environment.yml
conda activate brightpro-eval

cp .env.example .env       # fill OPENAI_API_KEY (judge + GPT-5-mini agent),
                           # HF_TOKEN (gated EmbeddingGemma), and optionally
                           # AZURE_OPENAI_* if going via Azure
chmod 600 .env

The data loader defaults to streaming yale-nlp/Bright-Pro from HuggingFace. For offline runs, point BRIGHT_PRO_DATA_ROOT at a local mirror of the dataset (layout: <root>/{examples,documents,aspects}/<task>.json).

Static retrieval

cd retrieval
python run.py --task biology --model rtriever-4b
# outputs/biology_rtriever-4b/score.json     <- {qid: {doc_id: score}}
# outputs/biology_rtriever-4b/results.json   <- standard NDCG/MAP/Recall/P/MRR at k ∈ {1,5,10,25,50,100}

--model can be any of the 13 retrievers above (see run.py --help for the exact CLI names). Standard metrics (NDCG@k, Recall@k, MAP@k, MRR) are written by run.py; aspect-aware metrics (α-nDCG@k and weighted A-Recall@k) are computed post-hoc:

python evaluation/alpha-ndcg-evaluation.py \
    --task biology --score_file outputs/biology_rtriever-4b/score.json \
    --k 25 --alpha 0.5

Agentic retrieval

The agentic protocol is run in two flavours over a fixed 175-query subset (agentic_sample_ids.json — 25 queries / task, seed=42):

Fixed-round (R ∈ {1,2,3}, top-5 per round):

cd agentic_retrieval
python search_agent/openai_fixed_turn.py \
    --searcher-type rtriever-4b \
    --tasks biology earth_science economics psychology robotics stackoverflow sustainable_living \
    --use-agentic-sample

Adaptive-round (agent decides when to stop):

python search_agent/openai_new.py \
    --searcher-type rtriever-4b \
    --tasks biology earth_science economics psychology robotics stackoverflow sustainable_living \
    --use-agentic-sample

Switch the OpenAI agent for the Qwen3.5 agent by replacing openai_*.py with qwen_new.py and supplying a vLLM endpoint via OPENAI_BASE_URL.

LLM-as-Judge

cd agentic_retrieval
python scripts_evaluation/judge.py \
    --runs-dir runs/gpt-5-mini-08-07 \
    --save-jsonl runs/gpt-5-mini-08-07/_judgements.jsonl \
    --deployment gpt-5

The judge prompt is the one used in the paper (the verbatim text is the JUDGE_INSTRUCTION constant at the top of judge.py). For each system answer the judge scores every reasoning aspect on {0, 0.5, 1} and assigns an overall answer score on a 1–5 Likert. Aspect coverage is aggregated using the paper's annotated aspect weights.

License

Released under the MIT License.