NorskHelsenett/eti-embedding-training-data-v2

ETI Embedding Training Data v2

576,708 Norwegian-language (anchor, positive, negative) triplets for contrastive fine-tuning of embedding, sparse-retrieval, and reranker models in the Norwegian health and welfare domain.

This dataset is a weighted, deduplicated merge of four upstream triplet datasets, designed to give a single high-quality training signal for two-stage RAG pipelines that retrieve over Norwegian public health information.

The model trained on it — NorskHelsenett/eti-embeddinggemma-v2 — reaches NDCG@10 0.8970 on a held-out test subset (vs. 0.6139 for the previous eti-embeddinggemma-v1 on the same protocol).

Schema

Column	Type	Description
anchor	string	The query — what a user would type into the search box. 10-500 chars.
positive	string	The gold document that answers the anchor. 100-8000 chars.
negative	string	A hard negative — a document that looks relevant but isn't. 100-8000 chars.
source	string	Which upstream dataset this triplet came from (provenance tag).

How to use

from datasets import load_dataset

ds = load_dataset("NorskHelsenett/eti-embedding-training-data-v2", split="train")
print(ds[0])
# {
#   "anchor":   "Hva er symptomene på diabetes type 2?",
#   "positive": "Diabetes type 2 gir ofte tørste, hyppig vannlating ...",
#   "negative": "Diabetes type 1 oppstår oftest i barne- eller ungdomsår ...",
#   "source":   "eti-embedding-training-data-2048-v22-hardnegs:triplets",
# }

For a recommended training recipe see NorskHelsenett/eti-embeddinggemma-v2 — fine-tunes google/embeddinggemma-300m with CachedMultipleNegativesRankingLoss (scale=20.0, mini_batch_size=24) at lr=1.5e-5, batch=96 per GPU × 2 GPUs, 1 epoch, with per-column prompts "task: search result | query: " (anchor) and "title: none | text: " (positive/negative).

Provenance

This is a weighted merge of four upstream triplet datasets:

Source	Rows pre-dedup	Weight	What it contributes
thivy/eti-embedding-training-data-2048-v22-hardnegs (triplets config)	~115k	×5	Production-distribution queries: anchors generated by the rewriter LLM that runs in front of the embedding model in the final RAG pipeline, paired with corpus-mined hard negatives. Highest-fidelity signal.
thivy/eti-embedding-training-data-2048-triplets-v4	~110k	×3	Multi-style anchors (informal/formal/short/long) + domain-vocabulary paraphrase pairs + antipattern negatives + LLM-judge filter.
thivy/eti-embedding-training-data-2048-triplets-v3-cleaned	~95k	×1	Earlier semantically-chunked triplets with mined hard negatives.
thivy/eti-embedding-training-data-2048-v2-triplets-v21-cleaned	~90k	×1	Original triplet set, retained for diversity.

Merge protocol (see scripts/merge_v3.py in the data-creation repo):

1. Per source: drop extra columns, apply length and mojibake filters, dedup within source.
2. Cross-source dedup at weight=1 (a triplet appearing in two sources is kept once, in input order).
3. Apply integer weights by row repetition.
4. Shuffle with seed=42.

Filters

• Length: anchor 10-500 chars; positive and negative 100-8000 chars.
• Mojibake: regex-filter UTF-8-as-Latin-1 garbage (e.g. å for å, √∏ for ø) that survives from upstream scrapes.
• Dedup: exact-tuple match on (anchor, positive, negative).

Source data

All articles ultimately derive from NorskHelsenett/LOS_Document_classification_ETI — the LOS welfare-ontology corpus of Norwegian public-health and welfare articles. Anchors are either:

• LLM-generated (Azure OpenAI gpt-5.1/gpt-4.1) using the production rewriter prompt or v4 multi-style prompts.
• Mined from production search-query logs (in the v22 subset).

Negatives are either:

• Corpus-mined hard negatives (sentence-transformers mine_hard_negatives, ranks 10-100 of the LOS corpus) with LLM-judge filtering.
• Synthetic antipattern negatives that codify failure modes named in the rewriter spec.

Intended use

• Fine-tuning bi-encoder embedding models for Norwegian information retrieval — both fresh-start and adaptation from multilingual bases (EmbeddingGemma, multilingual-e5, granite-embedding-r2, BGE, etc.).
• Fine-tuning SPLADE sparse encoders for learned-sparse retrieval over Norwegian text (the anchor/positive/negative shape is the native input for SparseMultipleNegativesRankingLoss + SpladeLoss).
• Fine-tuning cross-encoder rerankers — reshape each triplet into two labeled pairs (anchor, positive, 1) + (anchor, negative, 0) for BinaryCrossEntropyLoss.

Statistics

• Total rows: 576,708 (after dedup and weighting)
• Language: Norwegian (Bokmål primary, some Nynorsk)
• Domain: Health, welfare, public services
• Average anchor length: ~80 chars
• Average document length: ~2,000 chars
• Disk size: ~3.9 GB uncompressed / ~2.0 GB on the hub

Limitations

• All training pairs are Norwegian; the dataset is not multilingual.
• Anchors are LLM-generated, not from genuine human search queries (except the v22 subset). Models trained on this will perform best on inputs that look like LLM-rewritten queries.
• Negatives are corpus-internal — performance on out-of-domain distractors is not measured here.
• Mojibake filter is regex-based and catches the common patterns but won't catch all encoding bugs from upstream.

Citation

If you use this dataset, please cite the NorskHelsenett ETI work:

@misc{eti-embedding-training-data-v2,
    title  = {ETI Embedding Training Data v2: Norwegian Health Triplets for Two-Stage RAG},
    author = {Norsk Helsenett},
    year   = {2026},
    url    = {https://huggingface.co/datasets/NorskHelsenett/eti-embedding-training-data-v2}
}

License

CC-BY-4.0. Source LOS data is published by Norsk Helsenett SF — check the source dataset for any redistribution constraints on the underlying articles.