NorskHelsenett/eti-embedding-training-data-2048-v3

Version note (v3): third generation of the ETI dense training data. Reuses the questions of NorskHelsenett/eti-embedding-training-data-v2 (which trained eti-embeddinggemma-v2), but re-maps every question to its article URL, re-cuts positives at a 2048-token window, re-mines hard negatives with granite + reranker validation (pos_score/neg_score), adds the keyword half, and filters junk anchors. Trained NorskHelsenett/eti-embeddinggemma-v3.

eti-embedding-training-data-2048

Triplet dataset (anchor, positive, negative) for fine-tuning dense retrievers on Norwegian health and welfare information (helsenorge.no, NAV, Statped, municipal health services, and other Norwegian health authorities). Built for ibm-granite/granite-embedding-311m-multilingual-r2 with a 4096-token document window; also the source dataset for 2048-token rechunked variants (e.g. for google/embeddinggemma-300m).

Query styles match production two-stage RAG: every user question is rewritten into a keyword base_query before it reaches the retriever, so the dataset contains both natural questions and keyword queries.

Note: this repo previously hosted 78,888 anchor–positive pairs (2048-token chunks). That version remains available in the git history (revisions before 2026-08-25). Current content was formerly published as thivy/eti-embedding-training-data-4096-granite-v2.

Columns

column	description
anchor	query (natural NB or keyword)
positive	gold article at <=4096 granite tokens (full_text or granite best-chunk)
negative	reranker-validated hard-negative article at <=4096 granite tokens
positive_url / negative_url	source article URLs
pos_score / neg_score	domain reranker raw logits (query vs article summary_facts)
score_diff	pos_score - neg_score (>= 1.0 for every row)
query_style	natural or keyword

Build

1. Corpus — 2,019 articles from ES eti-summaries-v1 (1,970 full_text) + the 49 eti-sources sources-only articles (negative pool). Articles <=4096 granite tokens (84%) are used verbatim; >4096 (16%) use granite best-chunk (rank the article's eti-sources chunks by cosine to the anchor, greedily pack <=4096 tokens).
2. Positives — natural half: v2-triplet positives mapped to their article; keyword half: SPLADE-v14-distill summary_facts positives mapped to their article.
3. Negatives — granite bi-encoder mines candidates (rank window 5-60, cosine < 0.97), a same-service filter drops topical near-duplicates (same service across kommuner / sources / Nynorsk), and a domain cross-encoder keeps only candidates

   = 1.0 logit below the gold. Weak positives (gold reranker logit < -4.0) are dropped; final score_diff > 1.0 enforced.

Stats

• 555,023 triples · 202,112 anchors · 2,003 positive articles · 2,019 negative articles
• query styles: natural (281,937) + keyword (273,086)
• mean 2.7 negatives/anchor · score_diff mean 8.37 (min 1.00)
• ~101 anchors/article — high redundancy: train with CachedMultipleNegativesRankingLoss
  • the no_duplicates batch sampler to avoid two anchors of the same article becoming false in-batch negatives.

Intended use

Dense fine-tuning with CachedMultipleNegativesRankingLoss (in-batch + supplied hard negative). score_diff is provided for an optional MarginMSE distillation variant.

Training / evaluation code: the HN-ETI-AI-EMBEDDING-DATA repo (dataset-creation/ + dense-training/).