README.md

metadata

title: CTFAlign Demo
emoji: 🔗
colorFrom: indigo
colorTo: green
sdk: gradio
sdk_version: 5.9.1
app_file: app.py
pinned: false
license: mit
short_description: Demo of document-level word alignment algorithms.
preload:
  - Qwen/Qwen3-Embedding-4B
  - jhu-clsp/mmBERT-base
  - sentence-transformers/LaBSE

CTFAlign · Document-level Word Alignment — demo

Interactive demo for CTFAlign: choose one of the models from the paper's experiments, paste a document pair, and see the predicted word alignments. Hover a word to highlight its aligned counterpart(s) in the other text.

• Model — experiment models with a document-level tuned layer: Qwen3-Embedding-4B (default), LaBSE, mmBERT-base.
• Method — ctfalign (default) or mdpalign.
• Mode — base SimAlign variant: argmax (default) or itermax.
• Language pair — used to pick the tuned encoder layer; leave on (auto) to use the modal-best layer for the model. Layer and k/w are auto-filled with the values actually used (from the model / language pair / method) and stay editable.
• Context window — small-context models (LaBSE, 512 tokens) warn when a document exceeds their limit; prefer a larger-context model for long documents.

Exporting Alignment Labels

Export alignment (accordion under the viewer) exposes the same result the /align endpoint returns: the Pharaoh string (i-j …,), a copyable table of aligned pairs with their surface strings, the full JSON, and alignment.json / alignment.talp downloads.

Programmatically:

from gradio_client import Client

r = Client("miwytt/ctfalign-demo").predict(
    "mmBERT-base", "ctfalign", "argmax", "(auto)", None, None,
    source_text, target_text, "auto", api_name="/align",
)
r["pharaoh"]              # "0-0 1-1 2-4 ..."
r["alignment"]            # [{"s": 0, "t": 0, "source": "...", "target": "..."}, ...]
r["source"]["units"]      # the units the indices refer to
r["source"]["level"]      # "word" or "subword" (indices are meaningless without it)
r["config"]               # model / method / mode / layer / k / segmentation actually used
r["baseline_simalign"]    # the unconstrained SimAlign baseline, same shapes
r["warnings"]             # e.g. the text was chunked -> degraded quality

Run locally with pip install -r requirements.txt gradio && python app.py.