Palette_nlu_service — Persian Sales NLU

مدل تخصصی تشخیص Intent / Slot / Search-Query برای فروش فارسی.

• مدل پایه برای ترین: sbunlp/fabert
• معماری سرویس: encoder مشترک + سرهای Intent / Slot / Query
• Output schema: structured JSON matching the Palette INTENT_PROMPT

Intents (23)

PRODUCT_SEARCH, PRODUCT_PRICE, PRODUCT_DETAIL, PRODUCT_AVAILABILITY, PRODUCT_RECOMMENDATION, PRODUCT_COMPARE, PRODUCT_IMAGE, NEGOTIATION, ORDER_TRACKING, FAQ_*, GREETING, WELLNESS, THANKS, FAREWELL, HELP, OUT_OF_SCOPE

Files

Path	Description
checkpoints/best.pt	Trained weights
configs/model_config.yaml	Head sizes / pretrained name
label_maps/	intent2id / slot2id
tokenizer/	tokenizer assets

Load locally

import torch
from src.models.joint_bert import JointBERT
from src.utils.config import load_joint_bert_config

model = JointBERT(load_joint_bert_config("configs/model_config.yaml"))
ckpt = torch.load("checkpoints/best.pt", map_location="cpu", weights_only=False)
model.load_state_dict(ckpt["model_state_dict"], strict=False)
model.eval()

Serve API

See the companion repo Palette_nlu_service — python scripts/serve.py.