thivy/eti-embedding-training-data-2048-triplets-v4

ETI Embedding Training Data — v4 (5-style, LLM-judged triplets)

Norwegian (nb) retrieval triplets built from the NorskHelsenett/LOS_Document_classification_ETI corpus of public-service / welfare / health documents.

Each row is a triplet (anchor, positive, negative) plus three metadata columns (style, category, doc_url) you can use to filter, weight, or build curriculum stages.

76,408 triplets · 2,530 source documents · 38,629 distinct anchors.

Schema

field	type	description
anchor	string	Query in one of seven styles (see below).
positive	string	Chunk from the source document that should rank first.
negative	string	Cross-document hard negative — same topic family, wrong answer. Mined w/ FAISS, then LLM-judged.
style	string	One of tight_canonical, loose_canonical, vague_metaphor, mid_parent, long_parent, vocab_*.
category	string	Source LOS category (e.g. Helse og omsorg, Sosial og økonomisk trygghet).
doc_url	string	URL of the source document. Same doc_url never appears in both positive and negative.

Construction pipeline

Built by a five-stage pipeline run on Azure OpenAI:

Stage A — anchor generation (per chunk)
  └─ 5 styles × every chunk, generated by gpt-4.1
     tight_canonical   "Hjelpestønad ved langvarig sykdom"     ← rewriter-output-shaped
     loose_canonical   "Hva er hjelpestønad?"                  ← canonical noun + light scaffolding
     vague_metaphor    "Litt drahjelp når man står helt fast"  ← metaphor, no canonical noun
     mid_parent        "Vi sliter med å få endene til å møtes, finnes det noe vi kan søke?"
     long_parent       (3–5 sentence parent narrative)
Stage B — DOMAIN VOCAB paraphrase pairs (everyday ↔ canonical Norwegian welfare terms)
Stage C — Antipattern hard-negative descriptions (synthesised failure modes of the rewriter prompt)
Stage D — FAISS hard-negative mining (anchor → top-k chunks from a different doc_url)
Stage E — LLM-as-judge filter
  └─ gpt-5.1 evaluates every (anchor, positive, negative) — drops triplets where
     the model judges the negative more relevant than the positive (false negatives).
     ~87% of mined triplets survived.

Same-document chunks are excluded from the negative pool to prevent near-duplicate trivial pairs.

Why so many styles?

The target deployment is a two-stage RAG pipeline:

parent query  →  rewriter LLM  →  embedding model  →  retrieved docs

The rewriter prompt is itself under iterative development — its exact output distribution will shift over time. Training the embedding model on multiple anchor styles spanning the full pipeline (raw parent language → canonical rewriter output → metaphor) makes the embedding robust to those shifts, instead of overfitting to today's rewriter behaviour.

How v4 differs from v3

thivy/eti-embedding-training-data-2048-triplets-v3 was a single-purpose embedding-training set. v4 is a superset of v3's ideas designed to also work for reranker training and evaluation.

dimension	v3	v4
Anchor styles	2 (3 expert + 2 parent prompts per chunk)	5 distinct styles + 2 domain-vocab rows
Anchor LLM	gpt-4.1	gpt-4.1 (same — sufficient quality, fast)
Vocab paraphrase	—	Yes — explicit everyday ↔ canonical pairs
Anti-pattern negs	—	Yes — synthesised from the rewriter prompt's documented failure modes
Hard-negative mining	FAISS, cross-doc	FAISS, cross-doc
Same-doc filter	✓	✓
LLM-as-judge filter	—	Yes — gpt-5.1 verifies every triplet, drops false negatives (~13% removed)
Metadata columns	only anchor/positive/negative	adds style, category, doc_url for filtering / curriculum / eval splitting
Row count	43,322 raw → 36,870 cleaned	87,672 mined → 76,408 judged
Intended task	embedding fine-tuning	embedding fine-tuning + reranker training + retrieval eval

Recommended uses

1. Bi-encoder / embedding fine-tuning

Drop-in for CachedMultipleNegativesRankingLoss or TripletLoss in sentence-transformers. The five style buckets give the model a wide distribution of anchor surfaces; the same-doc filter on negatives prevents trivial-pair overfitting.

from sentence_transformers import SentenceTransformer, losses
from datasets import load_dataset

ds = load_dataset("thivy/eti-embedding-training-data-2048-triplets-v4", split="train")
ds = ds.select_columns(["anchor", "positive", "negative"])
loss = losses.CachedMultipleNegativesRankingLoss(model, mini_batch_size=24)

2. Cross-encoder / reranker training

Stage E (gpt-5.1 judge) gives every triplet a verified relevance ordering — positive > negative is not a heuristic, it's a strong-LLM verdict. That makes v4 well-suited for cross-encoder training, where clean pairwise labels matter more than volume.

Treat each row as two labelled pairs:

# pseudo-code
for row in ds:
    yield (row["anchor"], row["positive"]), 1.0
    yield (row["anchor"], row["negative"]), 0.0

For curriculum learning, increase difficulty by sorting on style: tight_canonical → loose_canonical → mid_parent → long_parent → vague_metaphor.

3. Retrieval evaluation

Use doc_url to build a clean held-out eval set without document leakage:

held_out_docs = {url for i, url in enumerate(ds["doc_url"]) if hash(url) % 10 == 0}
eval_set  = ds.filter(lambda x: x["doc_url"] in held_out_docs)
train_set = ds.filter(lambda x: x["doc_url"] not in held_out_docs)

Style distribution

style	rows	% of total
long_parent	16,726	21.9%
mid_parent	16,530	21.6%
vague_metaphor	14,920	19.5%
loose_canonical	14,646	19.2%
tight_canonical	13,003	17.0%
vocab_everyday	318	0.4%
vocab_canonical	265	0.3%

Source data

Derived from NorskHelsenett/LOS_Document_classification_ETI — a Norwegian public-service corpus covering welfare, health, education, work and culture services. Chunks were produced with LlamaIndex SemanticSplitterNodeParser (fallback SentenceSplitter at 2,048 tokens).

Caveats

• All anchors are LLM-generated; they are stylistically diverse but they are not real user queries. Pair with a small real-query eval set before claiming production gains.
• Judge filtering relies on gpt-5.1; residual judge errors are inherited by the dataset.
• The vocab_* styles are rare (~0.8% combined). Up-weight them if you specifically want to bake in the welfare-vocabulary mappings.
• Hard negatives are FAISS-mined from a single encoder snapshot; a bootstrapped second round of mining against the trained model would likely surface more discriminative negatives.

Citation

@misc{eti-embedding-training-data-v4,
  author       = {thivy},
  title        = {ETI Embedding Training Data — v4 (5-style, LLM-judged)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/thivy/eti-embedding-training-data-2048-triplets-v4}}
}