Multilingual_Intent_Classifier_checkpoint_v1

PyTorch checkpoint behind blue-machines/Multilingual_Intent_classifier_v1. Use this repo to keep fine-tuning. The Hub ONNX file is the INT8 deploy build of these weights, with the linear intent head left in FP32.

Model

8-layer Gemma 3 text stack, hidden size 320, about 100 million parameters. A mean pool over the encoder states feeds a linear intent head.

Intents

provide_info, affirm, deny, correction, question, clarify_request, unclear

Languages

Hindi, Tamil, Telugu, Kannada, Marathi, Malayalam, Bengali, Gujarati, Odia, and English. Utterances may be in the native script, romanized, or code-mixed.

Load

AutoModel cannot construct this student. Load Gemma3Intent8LStudent from the training code, then the weights in this repo:

from pathlib import Path

from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from transformers import AutoConfig, AutoTokenizer

from train_intent_multilang_indic_gemma270m_8L_distill import Gemma3Intent8LStudent

repo = "blue-machines/Multilingual_Intent_Classifier_checkpoint_v1"
local = Path(snapshot_download(repo))
config = AutoConfig.from_pretrained(local)
tokenizer = AutoTokenizer.from_pretrained(local)
model = Gemma3Intent8LStudent(config)
missing, unexpected = model.load_state_dict(
    load_file(local / "model.safetensors"), strict=False
)