Tiangong-LCA/LCA-Retrieval-Benchmark

LCA Retrieval Benchmark

A process-dataset retrieval benchmark for life-cycle assessment, released by Tiangong-LCA. It contains 100 queries from 42 source groups, 15,671 candidate documents, and 14,182 judged query–document relationships with probabilistic reference relevance weights.

Related model: LCA-Nemotron-Embed-1B. Training resource: LCA Embedding Training.

Task

Given a query, retrieve suitable LCI process records from the entire 15,671-document corpus, using the frozen text in corpus.jsonl. The query combines five fields:

Field	Meaning
P	Product/function or process-selection need
B	Process role and system boundary
G	Geography
T	Technology
Y	Applicable time period

Each field specifies a hard constraint or preference. Optional context is appended. Use queries.text directly: it preserves the exact query string used for the reported scores.

Files and loading

Configuration / file	Contents
queries / queries.jsonl	100 query IDs, texts, structured elements, source groups and text hashes
corpus / corpus.jsonl	15,671 document IDs, texts, source databases and text hashes
qrels / qrels.jsonl	14,182 reference relevance weights; includes judged zero-weight entries
model_rankings / model-rankings.jsonl	Verified Top10 from the released LCA-Nemotron model
model-results.json	Expected aggregate metrics and exact model revision
evaluate.py	Standard-library scorer for saved rankings, not a model runner
checksums.json	SHA256 and size of every payload file

from datasets import load_dataset
repo = "Tiangong-LCA/LCA-Retrieval-Benchmark"
queries = load_dataset(repo, "queries", split="test", revision="main")
corpus = load_dataset(repo, "corpus", split="corpus", revision="main")
qrels = load_dataset(repo, "qrels", split="test", revision="main")

Pin revision to a commit SHA for reproducible evaluation. Files can also be read as ordinary UTF-8 JSONL without the Datasets library.

Scoring protocol

For each query, rank all corpus documents. Sort descending similarity, with document ID ascending as the deterministic tie-breaker. Submit ten distinct document IDs:

{"query_id":"CQ_TEST_001","document_ids":["<rank-1 ID>","<rank-2 ID>","..."]}

The example is schematic; real rows must have exactly ten valid IDs. Then run:

python evaluate.py --rankings model-rankings.jsonl --verify-model-results
python evaluate.py --rankings /path/to/your-top10.jsonl

Let p(q,d) be the frozen relevance weight; weights lie in [0,1] and sum to 5 for each query. For K=5 and 10:

Coverage@K = sum(p(q, d_i), i=1..K) / 5
nDCG@K = sum(p(q, d_i) / log2(i+1), i=1..K)
         / sum(1 / log2(i+1), i=1..min(K,5))

Aggregate scores are arithmetic means over all 100 queries, not source-group means. This is fixed five-reference normalization, not standard exhaustive graded-qrels IDCG. No exponential gain transformation is used. Coverage is probability-weighted reference coverage, not accuracy and not exhaustive corpus recall.

Unjudged retrieved documents receive zero for this frozen-pool score, but are not proven irrelevant. The scorer separately reports their counts. Keep judged zero entries distinct from absent/unjudged entries. A reference weight is a model-assisted top-five selection probability, not an independently calibrated truth probability.

Reference model result

Pure embedding retrieval, no online reranking:

Model	nDCG@5	nDCG@10	Coverage@5	Coverage@10
LCA-Nemotron-Embed-1B	0.37277	0.44037	0.33689	0.45895

The included rankings reproduce these aggregates with the supplied scorer. Re-running retrieval also requires the model's documented inference environment, prompts, mean pooling, normalization, and 512/2048 query/document token limits. Different precision, tokenization or document serialization may alter nearest-neighbor ties.

Sources and limitations

Queries describe process-selection needs derived from applied-LCA source material; source groups are preserved for clustered analysis.

Process-document source	Records
BAFU 2025 v2	11,747
Tiangong LCA records	1,893
USLCI / Federal LCA Commons	1,341
ELCD 3.2	608
PlasticsEurope Eco-profiles	49
worldsteel LCI data	33

BAFU attribution: Life Cycle Inventory database of the Swiss Federal Administration, BAFU:2025. Corpus texts are processed retrieval representations, not replacements for the original inventory databases. Document IDs preserve source/version identity.

Labels are model-assisted and candidate judgments are incomplete. This is an exploratory benchmark, not a blind holdout. Its size and scenario selection limit generalization. Estimate uncertainty by resampling the 42 source groups. Use this repository's corpus and keep benchmark queries and labels separate from training.

Citation

@misc{tiangong_lca_retrieval_benchmark_2026,
  title = {LCA Retrieval Benchmark},
  author = {{Tiangong-LCA}},
  year = {2026},
  url = {https://huggingface.co/datasets/Tiangong-LCA/LCA-Retrieval-Benchmark}
}