---
title: laya-fa-support
emoji: 🎯
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.29.0
app_file: app.py
python_version: "3.12"
short_description: Persian support triage — department + probabilities
license: apache-2.0
---

# laya-fa-support

Persian customer-support triage with [`tarfandoon/laya-fa-support`](https://huggingface.co/tarfandoon/laya-fa-support):
a message goes in, and in one forward pass the model decides which department should handle it
(`billing`, `technical`, `cancel`, `other`) with a calibrated probability for every option.
It generates no text, so there is nothing to parse and nothing to hallucinate.

- Fine-tuned from Laya `multilingual` (mmBERT-base, 322M params), `top` mode finetune of the
  decision head plus the last 4 encoder layers.
- Held-out [`persian_fa` benchmark](https://github.com/alipyth/laya-persian-benchmark) (64 unseen
  Persian messages, unseen prompts): **79.7%** accuracy (base model: 56.2%), macro-F1 **0.795**,
  ECE **0.103**.
- Handles formal and colloquial Persian, Finglish, and Persian mixed with English; Arabic ي/ك
  and missing half-spaces are normalised by the tokenizer.
- The confidence gate ("route automatically" vs "send to a human") is plain code — pick the
  threshold on your own labelled data. False `cancel` decisions are the costly ones, since
  they would close an account.

The examples are real benchmark messages (MIT licensed) the model never saw during training.