nishaantshah/nepglish-nlu-v3

NepGlish-NLU v3 — superseded by v4

Use [nishaantshah/nepglish-nlu-v4](https://huggingface.co/nishaantshah/nepglish-nlu-v4) instead. v3 was trained on a dataset whose frequency and account_type labels were partly random (a generator bug, since repaired), and with a loss that never trained the end-of-turn token.

Corrected figures for v3 (held-out test set, n = 162, measured 2026-09-14 by the project's evaluation harness with the training prompt): intent accuracy 99.38 % (161/162), intent macro-F1 98.29 %, slot micro-F1 97.0 % — frequency slot 0.72. An earlier version of this card listed a macro-F1 of 0.989, 267 test samples and a frequency-slot F1 of 0.78; none of those had a supporting measurement.

Training system prompt: Extract the intent and slots from the following NepGlish financial query.