Tiangong-LCA/LCA-Embedding-Training

LCA Embedding Training

Graded query–document supervision for pure embedding retrieval of life-cycle inventory (LCI) process datasets, released by Tiangong-LCA.

This snapshot contains the final-stage training data used for LCA-Nemotron-Embed-1B. The separate evaluation resource is LCA Retrieval Benchmark.

Contents

Configuration	Rows	Meaning
queries	1,256	Unique training queries, with Chinese companion text and scenario metadata
relations	12,605	12,408 primary relations and 197 additional-positive relations
corpus	6,228	Documents referenced by supervision or query anchors
sampling	2,048	Training draw inventory for one epoch; repeated queries are intentional
retention	11,848	Frozen teacher cosine scores for 1,200 unchanged training queries

The relation grade distribution is: grade 4 = 2,486; grade 3 = 549; grade 2 = 470; grade 1 = 8,262; grade 0 = 838.

Loading

from datasets import load_dataset

repo = "Tiangong-LCA/LCA-Embedding-Training"
# Replace main with the dataset commit SHA for reproducible work.
queries = load_dataset(repo, "queries", split="train", revision="main")
relations = load_dataset(repo, "relations", split="train", revision="main")
corpus = load_dataset(repo, "corpus", split="train", revision="main")

documents = {r["document_id"]: r["text"] for r in corpus}
questions = {r["query_id"]: r["text"] for r in queries}
example = relations[0]
print(questions[example["query_id"]])
print(documents[example["document_id"]])
print(example["grade"], example["label"])

The files are UTF-8 JSONL and can also be read with Python's standard json module. checksums.json fixes every published payload file; hashes cover uncompressed bytes.

Schema and label semantics

• queries: query_id, English text, companion text_zh, scenario_group, retrieval_style, anchor_document_id, and source_family. Training consumed text, not both languages as separate examples. An anchor records query provenance; its presence alone does not create a positive label.
• corpus: stable document_id, exact text, source_database, UTF-8 text_sha256.
• relations: joins by query/document IDs. label is direct, acceptable, hard_negative, or negative; grade preserves the numeric supervision. Grades 3/4 are direct positives at different preference levels; 2 is acceptable; 1 is a hard negative; 0 is a negative. Higher positive grade is preferred; equal grades do not imply a strict order. rank is retained annotation metadata. role distinguishes primary from additional_positive; position preserves ordering within each role. Explanations, evidence and explicit preferences are included.
• sampling: draw_id, query_id, role, and position. This is the pre-shuffle draw inventory, not a timestamped optimizer-step exposure log.
• retention: teacher cosine similarities, not relevance probabilities, used to preserve pairwise margins for unchanged queries. These are not new relevance labels.

Sampling retains every primary candidate. At most two additional positives per query are exposed in an epoch: positions 0–1, 2–3, then 4–5; if an epoch slice is empty, positions 0–1 are reused. The graded objective separates direct/acceptable positives from negatives and orders unequal-grade positives; unlisted query–document pairs are unjudged, not negatives.

Source and construction

Process documents are processed retrieval-text representations of Tiangong LCA records, BAFU 2025 v2, ELCD 3.2, USLCI / Federal LCA Commons, PlasticsEurope Eco-profiles and worldsteel LCI data. The source_database and stable document ID retain origin information. For BAFU, source attribution is: Life Cycle Inventory database of the Swiss Federal Administration, BAFU:2025. These text representations are not the original inventory databases and do not replace their structured exchange data or methodological documentation.

Queries and relevance/preference annotations were constructed and reviewed with language-model assistance. Each relation includes its grade and supporting evidence.

Use and limitations

Use for domain embedding fine-tuning, graded retrieval experiments and supervision analysis.

This is a train-only resource. For independent evaluation, split by query and source family rather than individual relation rows. The linked benchmark has separate queries and its own corpus serialization.

Synthetic query style, model-label bias, restricted candidate pools and truncated document inputs limit transfer. The published model used query/document limits of 512/2048 tokens, including prefixes and special tokens; this dataset retains full stored text.

Citation

@misc{tiangong_lca_embedding_training_2026,
  title = {LCA Embedding Training},
  author = {{Tiangong-LCA}},
  year = {2026},
  url = {https://huggingface.co/datasets/Tiangong-LCA/LCA-Embedding-Training}
}