nlp-ui-small

Small (~9M param) classifiers that power nlp-ui's offline data-profiling pipeline: column field-role classification (16 roles) and dataset layout classification (5 layouts). Same question shapes as TypeSafe Jev (choice with probabilities + confidence), none of the API calls.

Checkpoints (subfolders, download only what you need)

Folder	Params	fp32	Best at
fieldrole/	~8.7M	~35 MB	column roles (16-way)
layout/	~10k	~40 KB	dataset layouts (5-way)

Each folder holds model.safetensors, model.onnx, plus vocab.json / config.json for the field-role encoder.

Use (ONNX, no torch needed)

import json, numpy as np, onnxruntime as ort
from tokenize import split_key  # see repo: python/features.py

vocab = json.load(open("fieldrole/vocab.json"))
sess = ort.InferenceSession("fieldrole/model.onnx", providers=["CPUExecutionProvider"])
# tokens: int64[1,12] key-token ids, kind: int64[1] value-kind id,
# numeric: float32[1,64] (48 ngram counts/8 + 12 flags + 4 scalars)
logits = sess.run(None, {"tokens": t, "kind": k, "numeric": n})[0]
probs = np.exp(logits) / np.exp(logits).sum()

Or serve behind HTTP and consume from nlp-ui via the remote-classifier adapter (same choice/probabilities/confidence contract as Jev).

Training

Transformer encoder (d=320, 6 layers) + scalar fusion head, trained with AdamW and label smoothing on the seeded synthetic corpus in python/data.py (tens of thousands of column profiles). See python/train_small.py. Held-out metrics are printed at train time.

Honest limits

• Synthetic-trained: profiles outside the generator's scalar ranges are the main failure mode; abstain below your confidence floor (minRoleConfidence).
• 16-option single questions only; split 50+ option sets coarse-to-fine.
• Refit temperature on your own labeled columns before trusting probabilities.