NL→SQL intent classifier

Fine-tuned bert-base-uncased that labels a natural-language data question with the kind of SQL it needs: aggregate, compare, count, filter, trend. It routes questions in the RAG-assisted natural-language-to-SQL system and picks the chart type.

Training data: 1,000 of the 2,000 domain-neutral questions in the repo's data/intent_dataset.csv (400 per intent, 14 domains); the other 1,000 are the validation split.
Validation accuracy: 100.0%. That split is templated like the training data, so it says little; on 150 hand-written hard questions the model scores 84.7%, and 75% on questions from an unseen SaaS domain (see the repo's README).

from transformers import pipeline
clf = pipeline("text-classification", model="dangkhoa241/nl2sql-intent-model")
clf("average billing amount by insurance provider")

Downloads last month: 21

Safetensors

Model size

0.1B params

Tensor type

F32

Model tree for dangkhoa241/nl2sql-intent-model

Base model

google-bert/bert-base-uncased

Quantized

(34)

this model