Intent classifier v1

blue-machines/Intent-classifier-v1 is a sentence-level 7-class intent classifier.

Base: microsoft/Multilingual-MiniLM-L12-H384
Deploy artifact: model.onnx — dynamic INT8 including embeddings (~113 MB)
Input: a single bare utterance (no [user] / [assistant] tags)
Max length: 96

Intent labels

id	label
0	provide_info
1	affirm
2	deny
3	correction
4	question
5	clarify_request
6	unclear

Files

File	Role
`model.onnx`	Deploy — INT8 ONNX (~113 MB; embeddings + MatMul quantized; GeGLU head FP32)
`model_int8.onnx`	Same INT8 graph (explicit name)
`model_fp32.onnx`	FP32 ONNX reference
`model.safetensors`	PyTorch weights (`checkpoint-2400`)
`tokenizer.json` / `tokenizer_config.json`	XLM-R / MiniLM tokenizer
`label_map.json`	label ↔ id
`config.json`	Transformer config
`export_metadata.json`	export provenance
`sample200_int8_eval.json`	200-sample INT8 eval snapshot

Metrics (approximate)

Split	Accuracy	Notes
Full test (train-time eval, FP32)	~0.837	checkpoint-2400
200-sample INT8 (seed=1) all	0.86	natural domain mix
200-sample INT8 Muthoot	0.8057553956834532
200-sample INT8 synthetic	0.9836065573770492

Hard-fail fine-tunes after this checkpoint degraded accuracy — do not use hardft weights.

Inference

import json
from pathlib import Path

import numpy as np
import onnxruntime as ort
from transformers import AutoTokenizer

MODEL_DIR = Path(".")  # or snapshot_download("blue-machines/Intent-classifier-v1")
MAX_LEN = 96
LABELS = [
    "provide_info", "affirm", "deny", "correction",
    "question", "clarify_request", "unclear",
]

tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.unk_token

session = ort.InferenceSession(
    str(MODEL_DIR / "model.onnx"),
    providers=["CPUExecutionProvider"],
)

def predict_intent(sentence: str) -> dict:
    enc = tokenizer(
        sentence,
        return_tensors="np",
        truncation=True,
        max_length=MAX_LEN,
        padding=True,
    )
    feed = {
        "input_ids": enc["input_ids"].astype(np.int64),
        "attention_mask": enc["attention_mask"].astype(np.int64),
    }
    logits = session.run(None, feed)[0][0]
    x = logits.astype(np.float64)
    x = x - x.max()
    probs = np.exp(x); probs = probs / probs.sum()
    pred_id = int(probs.argmax())
    return {
        "intent": LABELS[pred_id],
        "confidence": float(probs.max()),
        "probs": {LABELS[i]: float(probs[i]) for i in range(7)},
    }

print(predict_intent("Haan, ye sahi hai. Proceed karo."))

Notes

Intent-only (no LID / LSD).
Trained on expanded sentence-direct data (synthetic + Muthoot), with Muthoot-upweighted continue-FT.
Prefer this INT8 deploy over the earlier MatMul-only INT8 export (~389 MB).

Downloads last month: -

Safetensors

Model size

0.1B params

Tensor type

F32