eeahugs/query-intent-setfit-v1-train

EEA Query Intent — setfit-v1 training data

The exact training mix used to train eeahugs/query-intent-setfit-v1, the multilingual query-intent classifier for the European Environment Agency website search. 81,404 rows, 28 languages, 5 intent classes.

License: CC-BY-NC-4.0 (non-commercial use only) — see "Provenance and licensing" below.

Format

JSON Lines. One object per line:

field	meaning
id	row id, <prefix>-<lang>-<intent>-NNNN
intent	one of question, exploratory, claim, retrieval, unknown
language	BCP-47-ish code, one of 28
text	the query text, casefolded (lowercase)
source_id	lineage id of the source row (null for authored rows)
source_type	synthetic_generated, synthetic_translated, etc.

Class distribution: question 24,598 · retrieval 21,102 · exploratory 15,964 · claim 15,946 · unknown 3,794.

Training recipe (for reproducibility)

SetFitModel on intfloat/multilingual-e5-small (backbone revision 614241f622f53c4eeff9890bdc4f31cfecc418b3), 2 epochs, batch 64, head LR 1e-2, backbone LR 1e-5, fixed seed 3 (selected from seeds 1-3 on a held-out calibration set). Five labels as above; the production route is binary — AI-eligible when P(question)+P(exploratory)+P(claim) >= 0.95.

This file is the complete training input. The held-out calibration set and the frozen acceptance exams are intentionally not published.

Integrity

SHA-256 of this file (first 16 hex chars): 1e0e75eb65749835

Provenance and licensing

stratum	rows (approx)	source
A	~50,700	in-house Gemma 31B native generation, 17 languages (AI-generated)
B	~16,800	GPT-5.6-luna native generation, 6 languages: cs, el, et, hu, lt, lv (declared: OpenAI model outputs, published with the owner's decision to include)
C	~5,800	NLLB-200-1.3B machine translation (CC-BY-NC-4.0), sl + sv
D	~4,500	legacy English-anchor translations, 11 languages (NLLB / opus-mt; the Maltese model is CC-BY-SA-4.0)
E	~4,700	hand-authored English short-query bank, NLLB/opus-translated

Because the dataset contains machine-translation outputs from NLLB-200 (CC-BY-NC-4.0) and opus-mt-en-mt (CC-BY-SA-4.0), this dataset is published under CC-BY-NC-4.0: attribution required, non-commercial use only, share-alike for the opus-derived Maltese rows. The companion model weights are published under MIT by the owner's decision; the upstream disclosures in the model card apply to any reuse of this data.