LLM-Specific Utility Benchmark (SpecUBench)

SpecUBench is a benchmark for measuring the LLM-specific utility of retrieved passages in retrieval-augmented generation (RAG). Instead of assuming that a "relevant" passage is equally useful to every reader, UtilityBench labels how useful each retrieved passage is for a specific LLM — i.e., how much the passage actually helps that model produce the correct answer.

The benchmark is built on six widely-used open-domain QA / retrieval datasets and provides, for several LLMs, the per-passage golden utility labels together with the underlying queries, retrieved passages, ground-truth answers, and evaluation scripts.

• Original code & methodology: github.com/Trustworthy-Information-Access/LLM_specific_utility
• Source dataset on ModelScope: hengranzhang/LLM-Specific-Utility-Benchmark

---

Overview

For a query and a candidate passage, the utility of the passage for an LLM is derived from the performance difference of the model when it answers with the passage versus without the passage:

1. Generate the model's answer with the retrieved passages;
2. Generate the model's answer without any passage (query-only);
3. Compute the exact-match / F1 difference between the two to obtain the passage-level golden utility signal.

This produces a per-model, per-query utility label for each retrieved passage, which can then be used to train or evaluate utility-aware retrievers and rerankers.

---

Source Datasets

The benchmark uses six datasets:

Dataset	Type	Source
Natural Questions (NQ)	open-domain QA	KILT / DPR
HotpotQA	multi-hop QA	KILT / DPR
TriviaQA	open-domain QA	KILT / DPR
FEVER	fact verification	KILT / DPR
MS MARCO	passage retrieval / QA	Microsoft MS MARCO
2WikiMultiHopQA (2WikiQA)	multi-hop QA	official 2WikiQA

For each dataset, source_datasets/<dataset>/ provides the queries, ground-truth answers, and the top-200 passages retrieved by a dense retriever (BGE, via Tevatron).

---

Evaluated LLMs

Golden utility labels are provided for the following models:

• Qwen3-8B
• Qwen3-14B
• Qwen3-32B
• Llama-3.1-8B-Instruct

---

Repository Structure

UtilityBench/
├── source_datasets/          # queries, answers, retrieved top-200 passages, qrels
│   ├── 2wikiqa/
│   ├── fever/
│   ├── hotpotqa/
│   ├── msmarco/
│   ├── nq/
│   └── triviaqa/
├── Golden_utility_passage/   # per-model golden utility passage labels
│   ├── 2wikiqa/
│   ├── fever/
│   ├── hotpotqa/
│   ├── msmarco/
│   ├── nq/
│   └── triviaqa/
│       └── golden_utility_passages/
│           ├── {model}_utility_passages.jsonl      # passage-level utility labels
│           └── {model}_utility_passages_qrel.txt   # TREC-format qrel
└── evaluation/               # evaluation & significance-test scripts and results

source_datasets/

Raw inputs for each dataset, including the queries (dev.jsonl), ground-truth answers (gt_answer.jsonl), the top-200 retrieved passages per query (top_200_passages.jsonl), and relevance judgments (qrel.txt / dev_rank.txt). Auxiliary files such as union.jsonl / union_positive.jsonl combine retrieved passages with the human-annotated / positive passages.

Golden_utility_passage/

For each dataset × LLM combination, the passages whose utility has been labeled for that specific model. The JSONL files carry the per-passage utility signal, and the companion .txt files provide the same information in TREC qrel format for retrieval evaluation.

evaluation/

Scripts for ranking-based and set-based evaluation (ranking-evaluation.py, set-evaluation.py), exact-match / F1 computation (em_f1.py, fever_acc*.py), and statistical significance testing (significance_test_*.json, significant_test.py).

---

Usage

The golden utility labels can be used directly to train or evaluate utility-aware retrievers and rerankers. For the full labeling and evaluation pipeline, refer to the official repository, which documents:

• Golden utility labeling — pointwise_performance_answerem.py, without_passage_answer.py, and the utility-computation scripts in evaluation/.
• LLM-specific utility judgment methods — src/answer_test.sh.
• Evaluation — evaluation/ranking-evaluation.py and evaluation/set-evaluation.py.

---

License

This dataset is released under the Apache License 2.0. The underlying datasets (NQ, HotpotQA, TriviaQA, FEVER, MS MARCO, 2WikiQA) retain their own licenses; please refer to their respective sources.

---

References

• Repository: github.com/Trustworthy-Information-Access/LLM_specific_utility
• ModelScope source dataset: hengranzhang/LLM-Specific-Utility-Benchmark