README.md

---
title: Flan T5 Token Ner
emoji: 📚
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 5.23.3
app_file: app.py
pinned: false
license: mit
short_description: Classifies each token in the input text as LOC, ORG, PER, or
---

# Flan-T5 Token Classifier (NER Demo)

This Huggingface Space is a Gradio demo for the model [`pepegiallo/flan-t5-base_ner`](https://huggingface.co/pepegiallo/flan-t5-base_ner). It performs **token-level Named Entity Recognition (NER)** using a Flan-T5 encoder-based architecture.

---

## 🔍 What does this demo do?

You can enter any sentence, and the app will:

1. Split the sentence into tokens (words and punctuation)
2. For each token:
    - Mark it with `<TSTART>` and `<TEND>` in the context of the sentence
    - Send it through the model with the prompt: `classify token in: <wrapped sentence>`
3. Predict one of the following labels for each token:
    - `PER` — Person
    - `ORG` — Organization
    - `LOC` — Location
    - `O` — Not an entity

---

## 🧠 Example

Input:
```
Max Mustermann works at Microsoft and lives in Berlin.
```

Output:
```
Max         -> PER
Mustermann  -> PER
Microsoft   -> ORG
Berlin      -> LOC
```

---

## 📦 Model Details

- **Base model:** `google/flan-t5-base` (encoder only)
- **Fine-tuned on:** WikiANN, open-pii-masking-500k, and custom samples
- **Prompt-based classification** per token
- **Architecture:** T5 encoder + classification head

---

## 🚀 Try it out!

Type any sentence in English, German, French, Italian or Spanish, and the model will tag names, organizations, and locations.

For more details, check the full model card:
👉 [`pepegiallo/flan-t5-base_ner`](https://huggingface.co/pepegiallo/flan-t5-base_ner)