LCA-Nemotron-Embed-1B
A domain-adapted pure embedding model for retrieving life-cycle inventory (LCI) process datasets from natural-language LCA queries. Released by Tiangong-LCA. It produces 2048-dimensional, unit-normalized vectors. Retrieval uses cosine similarity / inner product.
The full checkpoint is included: no adapter merge or separate base download is needed.
中文说明
本模型用于 LCA/LCI 过程数据检索,将查询和过程文档编码为 2048 维向量。
提供完整微调权重,无需合并适配器或下载独立基础模型。
查询使用 query: 前缀、512 token 上限;文档使用 passage: 前缀、2048 token
上限,均包含前缀及特殊 token。下面的代码自动添加前缀,不要手动重复添加。
在 LCA Retrieval Benchmark
上的 nDCG@10 为 0.44037,Coverage@10 为 0.45895。
Model details
| Property | Value |
|---|---|
| Upstream model | NVIDIA Nemotron-3-Embed-1B-BF16 |
| Upstream snapshot used | c0c9fea93ea424587517f2c59e20db9f1d6bf615 |
| Architecture | Ministral3Model, bidirectional attention (is_causal=false), 16 layers |
| Pooling | Attention-mask mean pooling, including prompt tokens, followed by L2 normalization |
| Embedding dimension | 2048 |
| Weight storage | Full FP32 safetensors; qualified inference uses BF16 on CUDA |
| Task-qualified lengths | Query 512; document 2048 tokens, truncation enabled |
| Languages | Chinese and English LCA retrieval |
| Release date | 2026-10-13 |
| License | OpenMDW-1.1; upstream attribution and third-party notices included |
Installation
Use a clean Python 3.12 environment. The tested environment is Linux x86_64, PyTorch 2.11.0+cu129, Transformers 5.4.0, Sentence Transformers 5.3.0, huggingface-hub 1.8.0 and safetensors 0.7.0.
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cu129
python -m pip install transformers==5.4.0 sentence-transformers==5.3.0 huggingface-hub==1.8.0 safetensors==0.7.0
Use a CUDA 12.9-compatible driver for this GPU environment. The helper also supports CPU FP32. Loading does not require remote code.
Quick start: Sentence Transformers
import torch
import torch.nn.functional as F
from sentence_transformers import SentenceTransformer
MODEL_ID = "Tiangong-LCA/LCA-Nemotron-Embed-1B"
MODEL_REVISION = "9bb397e200a199a05bd073f6074e4080e93b38dd"
model = SentenceTransformer(
MODEL_ID, revision=MODEL_REVISION, device="cuda:0", trust_remote_code=False,
model_kwargs={"dtype": torch.bfloat16, "attn_implementation": "sdpa"},
)
model.max_seq_length = 512
q = model.encode_query(
["Inventory dataset for electricity production from natural gas in China"],
batch_size=8, convert_to_tensor=True, normalize_embeddings=False,
)
model.max_seq_length = 2048
d = model.encode_document(
["Electricity production; natural gas combined cycle; geography: China; unit: kWh",
"Wheat grain production; geography: France; unit: kg"],
batch_size=8, convert_to_tensor=True, normalize_embeddings=False,
)
q = F.normalize(q.float(), dim=1)
d = F.normalize(d.float(), dim=1)
scores = q @ d.T
print(scores)
print(scores.argsort(dim=1, descending=True))
Supply original text without prefixes: encode_query and encode_document add
the saved prompts. Plain model.encode(texts) does not automatically choose a
query/document prompt. Do not use last-token pooling or causal attention.
Changing max_seq_length is shared mutable state; the included encode.py helper
serializes access when used concurrently.
Download and local batch use
hf download Tiangong-LCA/LCA-Nemotron-Embed-1B --local-dir ./lca-nemotron
cd lca-nemotron
python encode.py --model . --device cuda:0 --input-type query "中国天然气联合循环发电的单位电量清单"
python encode.py --model . --device cuda:0 --input-type document "Electricity production, natural gas, combined cycle; China; kWh"
The helper returns a JSON list of embeddings. In Python:
from encode import LCAEmbedder
encoder = LCAEmbedder(model=".", device="cuda:0", local_files_only=True)
queries = encoder.encode(["中国电网中压电力供应"], "query")
documents = encoder.encode(["Electricity, medium voltage; market; China; kWh"], "document")
print(queries @ documents.T)
Document formatting and indexing
Provide the process name and available description, geography, technology, reference product/unit, system boundary, time period and production/market role as readable text. Preserve meaningful fields consistently. This model does not parse raw TIDAS/ILCD JSON automatically; evaluation used a frozen corpus text serialization. The examples above are illustrative, not benchmark records. Changing serialization or truncation changes retrieval results. Encode every document with this checkpoint, store the 2048-dimensional normalized vectors, and search with inner product or cosine. Similarity is not a calibrated probability or an LCA suitability certification.
HTTP API example
After downloading the repository, start the local example with:
python serve.py --model . --device cuda:0 --port 8000
curl http://127.0.0.1:8000/v1/embeddings \
-H 'Content-Type: application/json' \
-d '{"model":"Tiangong-LCA/LCA-Nemotron-Embed-1B","input_type":"query","input":["中国天然气发电清单"]}'
Use "input_type":"document" for corpus documents. Output has data[i].embedding
and data[i].index. The API uses an OpenAI-style response and an explicit
input_type. The example binds to 127.0.0.1 and limits request size and batch count.
Datasets
- LCA Embedding Training: final-stage training snapshot with 1,256 queries, 12,605 graded query–document relations, document texts, sampling and retention supervision.
- LCA Retrieval Benchmark: evaluation queries, frozen corpus, probabilistic reference labels, this model's Top10 rankings and a standalone metric scorer.
Evaluation: LCA Retrieval Benchmark
The LCA Retrieval Benchmark uses a frozen full-query protocol: 100 tasks from 42 source groups, 15,671 corpus documents, pure embedding retrieval. Metric definitions and reproducible inputs are provided in the benchmark repository.
| Model | nDCG@5 | nDCG@10 | Coverage@10 |
|---|---|---|---|
| LCA-Nemotron-Embed-1B (this release) | 0.37277 | 0.44037 | 0.45895 |
| Qwen3-Embedding-4B | 0.38022 | 0.43944 | 0.44292 |
| Codestral Embed 2505 | 0.35223 | 0.43095 | 0.46140 |
| Qwen3-Embedding-8B | 0.35452 | 0.41328 | 0.42094 |
| Nemotron-3-Embed-1B (base) | 0.31692 | 0.37708 | 0.41268 |
| LCA-Qwen3-0.6B | 0.25161 | 0.31187 | 0.34093 |
| Qwen3-Embedding-0.6B | 0.18926 | 0.23031 | 0.25181 |
| GTE-large | 0.15506 | 0.19161 | 0.20031 |
| BGE-M3 | 0.14949 | 0.18814 | 0.21381 |
| all-mpnet-base-v2 | 0.14137 | 0.16945 | 0.17999 |
Compared with the base model, nDCG@10 rises by 0.06330 (about 16.79% relative). The linked benchmark includes this model's Top10 rankings and a standalone scorer. Its nDCG and Coverage use fixed five-reference normalization.
Intended use and limitations
- Candidate retrieval for LCA/LCI research and process-dataset matching; not text generation.
- Does not validate inventory quality, system-boundary compatibility, units or regulatory compliance.
- Similar-looking processes may differ materially in geography, production versus market, technology, energy input/output, or treatment versus avoided production.
- Long inputs are truncated under the qualified protocol; important late fields can be lost.
- Limited domain coverage and synthetic/model-assisted supervision can introduce annotation bias, domain overfitting and evaluation optimism.
- Benchmark gains do not guarantee improvements on other datasets.
Release integrity and licensing
release-manifest.json records file sizes and SHA-256 hashes for integrity checks.
Weights are stored in FP32; BF16 is selected at load time. validation.json records
inference checks. NVIDIA/Ministral notices are included in LICENSE, NOTICE
and THIRD_PARTY_NOTICES.md; CHANGES.md identifies downstream modifications.
Review these files and upstream terms before reuse. No endorsement by NVIDIA or
Mistral is implied. Additional wrapper code and documentation are provided under
OpenMDW-1.1. Third-party software retains its own licenses.
Citation and contact
@misc{tiangong2026lcanemotron,
title = {LCA-Nemotron-Embed-1B: Domain-Adapted Embeddings for Life-Cycle Inventory Retrieval},
author = {{Tiangong-LCA}},
year = {2026},
howpublished = {\url{https://huggingface.co/Tiangong-LCA/LCA-Nemotron-Embed-1B}}
}
Please also credit the upstream NVIDIA model. For feedback, open a discussion on this model repository or see Tiangong LCA.