Datasetpublic

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

ETI Embedding Training Data — Triplets with Hard Negatives This dataset contains 330,120 (anchor, positive, negative) triplets for training and fine-tuning Norwegian-language embedding models, particularly for health-related retrieval and RAG applications. How this dataset was created Source data The triplets were mined from the source dataset NorskHelsenett/eti-embedding-training-data-2048, which contains 78,888 anchor-positive pairs of Norwegian… See the full description on the dataset page: https://huggingface.co/datasets/NorskHelsenett/eti-embedding-training-data-2048-triplets.

sourceHugging Facecc-by-4.0updated 5mo agoView on Hugging Face

0likes60downloads

Dataset Card

ETI Embedding Training Data — Triplets with Hard Negatives

This dataset contains 330,120 (anchor, positive, negative) triplets for training and fine-tuning Norwegian-language embedding models, particularly for health-related retrieval and RAG applications.

How this dataset was created

Source data

The triplets were mined from the source dataset [NorskHelsenett/eti-embedding-training-data-2048](https://huggingface.co/datasets/NorskHelsenett/eti-embedding-training-data-2048), which contains 78,888 anchor-positive pairs of Norwegian health content. That dataset was generated using an LLM (Claude) to create questions (anchors) for passages (positives) from Norwegian public health sources (helsenorge.no, NAV, Statped, etc.).

Hard negatives mining

Hard negatives were mined using Sentence Transformers' `mine_hard_negatives()` with FAISS-based approximate nearest neighbor search. The process:

1.All anchors and positives from the source dataset were embedded
2.For each anchor-positive pair, FAISS was used to find semantically similar but incorrect passages
3.Filtering was applied to ensure negatives are "hard" (close but not too close to the anchor)
4.Each input pair produced multiple triplets with different hard negatives

Parameters used

Parameter	Value	Description
`embedding_model`	`NorskHelsenett/eti-embeddinggemma-v1`	Domain-specific Norwegian health embedding model used for similarity search
`num_negatives`	`5`	Target number of hard negatives per anchor-positive pair
`range_min`	`5`	Skip the 5 most similar candidates (too easy/near-duplicates)
`range_max`	`200`	Consider top 200 most similar candidates
`max_score`	`0.9`	Reject candidates with similarity > 0.9 to anchor
`absolute_margin`	`0.1`	Negative must be ≥ 0.1 less similar than positive
`relative_margin`	`0.1`	Negative similarity ≤ 0.9× positive similarity
`sampling_strategy`	`top`	Select the hardest negatives from the valid range
`output_format`	`triplet`	One (anchor, positive, negative) row per negative
`use_faiss`	`True`	FAISS for memory-efficient approximate nearest neighbor

Statistics

Metric	Value
Source pairs	78,888
Output triplets	330,120
Avg triplets per anchor	~4.2
Dataset size	~503 MB

Dataset format

Each row contains three text fields:

Column	Description	Example
`anchor`	A question in Norwegian	"Hva er noen tips for å gjøre leken mer lystbetont for barnet mitt?"
`positive`	The correct/relevant passage	A passage about supporting language learning through play
`negative`	A hard negative — related but incorrect passage	A passage about a different but topically similar health subject

Intended use

This dataset is designed for:

—Fine-tuning embedding models using triplet loss or similar contrastive objectives
—Training cross-encoders for re-ranking in Norwegian health RAG systems
—Improving retrieval quality by teaching models to distinguish between similar-but-different health topics

Training with Sentence Transformers

python

from sentence_transformers import SentenceTransformer, losses
from datasets import load_dataset

dataset = load_dataset("NorskHelsenett/eti-embedding-training-data-2048-triplets", split="train")

model = SentenceTransformer("your-base-model")
train_loss = losses.TripletLoss(model=model)

# See SBERT docs for full training setup:
# https://sbert.net/docs/training/overview.html

Source dataset

[NorskHelsenett/eti-embedding-training-data-2048](https://huggingface.co/datasets/NorskHelsenett/eti-embedding-training-data-2048) — 78,888 anchor-positive pairs generated from Norwegian health content using Claude for question generation and semantic chunking with a 2048-token limit.

Related models

—[NorskHelsenett/eti-embeddinggemma-v1](https://huggingface.co/NorskHelsenett/eti-embeddinggemma-v1) — An embedding model fine-tuned on the source anchor-positive pairs

Pipeline

This dataset was created using the HN-ETI-AI-EMBEDDING-DATA pipeline's --hard-negatives-only mode:

bash

uv run main.py \
  --hard-negatives-only NorskHelsenett/eti-embedding-training-data-2048 \
  --mine-hard-negatives \
  --hard-negatives-model NorskHelsenett/eti-embeddinggemma-v1 \
  --num-hard-negatives 5 \
  --hard-negatives-range-min 5 \
  --hard-negatives-range-max 200 \
  --hard-negatives-max-score 0.9 \
  --push-name NorskHelsenett/eti-embedding-training-data-2048-triplets

License

CC-BY-4.0 — Created by Norsk Helsenett SF. Source content is from Norwegian public sector websites.