---
title: Bengali Headline Annotation
emoji: 📰
colorFrom: green
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
---

# Bengali Headline Annotation

Human pairwise comparison of model-generated headlines vs the gold reference.

## Setup (Space owner)

Set the following **Repository secrets** in Space → Settings:

- `HF_TOKEN` — write-scoped token; used to push annotations to the private dataset.
- `ANNOTATOR_PASSWORD` — shared password handed to annotators.

## Annotator instructions

1. Enter your name (single word, lowercase preferred — we auto-lowercase) and the password you were given.
2. For each article, read the news text (and image if shown), then pick which headline is better: **A**, **B**, or **Tie**.
3. Optional comment for each rating.
4. Progress is saved per name; you can close the tab and resume any time.

## Source code

- Builder, deploy script, and analyzer: <repo>/train-headline-gen/annotate/ and <repo>/train-headline-gen/analyze.py