lca-qwen3-embedding

Domain embedding model for lifecycle assessment (LCA) retrieval. It encodes sentences and short passages into 1024‑d L2-normalized embeddings for semantic search, similarity scoring, and clustering.

Background

Generic embedding models work well in open domains, but professional LCA retrieval often involves long, structured records (e.g., geography/technology/time fields) and domain-specific terminology. This model is trained to better align embeddings with LCA retrieval queries and documents.

Results (our evaluation setup)

On an internal evaluation derived from TianGong LCA records (converted from the Tidas structured format into retrieval-friendly text), this model improved over the base Qwen3-Embedding-0.6B on both ranking quality and tail coverage:

• vs base Qwen3-Embedding-0.6B: NDCG@10 +31.2%, Recall@10 +25.7%, MRR@10 +33.5%, Recall@100 +11.5%

Evaluation scale (this experiment):

• Train: 17,037 query-doc pairs
• Eval: 1,893 queries / 3,786 corpus docs / 1,893 qrels

Model comparisons

Key metrics (@10):

Model	NDCG@10	Recall@10	MRR@10	MAP@10
Qwen3-Embedding-0.6B (base)	0.5808	0.7200	0.5367	0.5367
lca-qwen3-embedding (this model)	0.7623	0.9049	0.7163	0.7163
codestral-embed-2505	0.6628	0.8045	0.6180	0.6180
qwen3-embedding-8b	0.5905	0.7369	0.5442	0.5442
qwen3-embedding-4b	0.5836	0.7290	0.5377	0.5377
bge-m3	0.5839	0.7264	0.5388	0.5388

Tail coverage (@100):

Model	NDCG@100	Recall@100
Qwen3-Embedding-0.6B (base)	0.6171	0.8922
lca-qwen3-embedding (this model)	0.7826	0.9947
codestral-embed-2505	0.6872	0.9171
qwen3-embedding-8b	0.6258	0.9033
qwen3-embedding-4b	0.6164	0.8822
bge-m3	0.6156	0.8743

Protocol note: embeddings are L2-normalized; retrieval uses inner product (equivalent to cosine similarity) with top-100 candidates.

Model details (from the exported config)

• Backbone: Qwen3 (model_type=qwen3; config architecture Qwen3ForCausalLM), hidden_size=1024, num_hidden_layers=28
• Max sequence length: 1024
• Embedding dimension: 1024
• Pooling: last-token pooling (pooling_mode_lasttoken=true, include_prompt=true)
• Normalization: L2 normalize
• Similarity: cosine
• Prompts: a query prompt is defined; the document prompt is empty

Module stack:

Transformer -> Pooling(last_token, include_prompt=true) -> Normalize

Usage (SentenceTransformers)

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("BIaoo/lca-qwen3-embedding")  # replace with your HF repo id if forked/renamed

Retrieval example (encode queries and documents separately; apply the built-in query prompt):

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("BIaoo/lca-qwen3-embedding")  # replace with your HF repo id if forked/renamed

queries = ["wood residue gasification heat recovery"]
docs = ["Report describing small-scale biomass CHP units used for district heating."]

q = model.encode(queries, prompt_name="query", normalize_embeddings=True)
d = model.encode(docs, normalize_embeddings=True)
scores = q @ d.T  # cosine similarity (because normalized)
print(scores)

Notes:

• Use prompt_name="query" to apply the query instruction prefix from config_sentence_transformers.json.
• The document-side prompt is empty; encoding documents with encode(docs, ...) is typically sufficient.

Intended use

• Semantic search and reranking for LCA process/flow descriptions and metadata-rich technical text
• Similarity scoring for deduplication / clustering of LCA-related passages

Limitations

• Trained and evaluated primarily on English technical/LCA text; performance may degrade in other languages or domains.
• Evaluation numbers are from a specific internal setup; validate on your own data before production use.

Files

• config.json: Qwen3 model config
• config_sentence_transformers.json, modules.json, sentence_bert_config.json: SentenceTransformers configs (prompts, modules, max length)
• model.safetensors: weights
• tokenizer.*, vocab.json, merges.txt: tokenizer assets
• 1_Pooling/, 2_Normalize/: pooling / normalization modules