e5-kz-skills

Bi-encoder that maps skill phrases written by employers in Kazakhstan (Enbek, hh.kz, hh.ru; Kazakh, Russian, English) to ESCO skills. Fine-tuned from intfloat/multilingual-e5-base with cached multiple-negatives ranking loss and mined hard negatives on 28599 pairs (employer phrasing equal to ESCO labels, confident ESA matches, ru/kk ESCO labels), split by skill. Use the e5 prefixes: query: for phrases, passage: for "label. description" of ESCO skills.

encoder recall@1 recall@10 MRR (dev, all ESCO skills)
base 0.6787 0.8533 0.7396
fine-tuned 0.8643 0.961 0.9005

No-match threshold (cosine): 0.3468 (dev precision 0.8669, recall 0.8643).

Limitations. Dev anchors are ESCO labels and employer phrases equal to labels; the expert-marked gold set of unmatched Kazakhstan phrasing (gold.csv) is the honest estimate. ESCO has no skills for some local notions (1ะก configurations, Kazakhstan regulations): those should stay unmatched.

Downloads last month
34
Safetensors
Model size
0.3B params
Tensor type
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Fralet/e5-kz-skills

Finetuned
(194)
this model