Batwan Predictive Text

Batwan Predictive Text

Suggests the rest of a word, or the next one, as you type in a Hiligaynon-English dictionary search box, using a small fine-tuned Gemma model. A demo for language learners and developers.

Made by alsersugasawa on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: This Space is a search box that predicts the rest of what you type, built for the Batwan Hiligaynon-English dictionary. It offers the rest of the current word or, after a space, the next word. The best guess appears as grey text after your cursor, and other guesses sit under the box.

Who it suits: It suits learners of Hiligaynon and the developers of dictionary or search sites who want to see predictive text on a small language. The README calls it a demo and rules out commercial use.

Industry | Who uses it | How they use it | Example
Education | Hiligaynon learners | Get help finishing words in a dictionary search | Type the start of a word and read the guesses
Software | Search developers | Study a predictive search box backed by a small model | Call the suggest endpoint with a short query
Language tech | Researchers | Look at a fine-tuned model on a dictionary dataset | Read how a suggestion is scored
Publishing | Dictionary editors | Judge whether guesses fit the dictionary's own words | Compare guesses with the headwords

Anyone who wants predictive text for a commercial product can skip it, since the README says it is only a showcase.

Where to use it: The demo runs in three places.

• A web browser, on the page of the Space at Hugging Face
• Windows, macOS or Linux, as a local copy run in Docker
• Another web page or script, through the Space's suggest endpoint

Use case | Who | What they get back
Word completion | Learners | The rest of the word you are typing
Next-word guess | Learners | A likely next word after you type a space
Suggestion API | Developers | A list of suggestions with a score for each

How to set it up: The online page needs nothing installed. A local copy needs the following.

• Docker
• Python with pip, to install the Hugging Face hub package
• A connection that can fetch the model, about 3 GB, when the container starts

1. Install the Hugging Face hub package and download a copy of the Space.
2. Move into the new folder and build the Docker image.
3. Run the container with port 7860 published.

pip install -U huggingface_hub
hf download alsersugasawa/batwan-predictive-text --repo-type space --local-dir batwan-predictive-text
cd batwan-predictive-text
docker build -t batwan-predictive-text .
docker run -it -p 7860:7860 batwan-predictive-text

How to use it: Type into the search box. The best guess shows as grey text, and pressing Tab takes it. The other guesses appear under the box. After a space, the box guesses the next word.

Tool | What it does
Search box | Shows the best guess in grey and the others below it
Tab key | Accepts the grey suggestion
Suggest endpoint | Takes q, lang, limit and words and returns suggestions with scores
q | The text typed so far, up to 100 characters, with a final space asking for the next word
lang | any, hil or en, the dictionary's search directions
limit | How many suggestions come back, 1 to 8, with 3 as the default
words | Extra words a suggestion may run on for, 0 to 3, with 0 as the default

Example: You type maayong a into the box. The grey text completes it to maayong aga, and a few other guesses show under the box, each one ranked by the model.

For newcomers and experts: If you are new to this, think of the suggestions on a phone keyboard. Type a few letters of a Hiligaynon word and watch the grey text, then press Tab to accept it.

If you know the field, the model is a LoRA fine-tune of Gemma 4 E2B, at 3 GB in Q4_K_M. It trained on search queries built from the dictionary's words, their English, phrases and AI-mode questions. A suggestion takes three steps. Token healing finds the word being typed, the three likeliest paths branch one token further, and the best three finish the word. Scores use the model's own top-k probabilities, and each suggestion is rated as a fraction of everything the search turned up.

You type in the search box
Token healing finds the current word
Three likeliest paths branch out
The best three finish the word
Guesses show in grey and below

Good to know: Suggestions take about two seconds on the free tier, because the model writes only a handful of tokens a second on two CPU cores. If you keep typing, the older pending request is dropped.

• License: Apache 2
• Cost to run: none named
• Data it reaches: your typing goes to the Space's own service, and the model weights and a training JSON file are fetched when it starts
• Not for: commercial use, since the README calls it a demo

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: docker
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/alsersugasawa/batwan-predictive-text
• Files: https://huggingface.co/spaces/alsersugasawa/batwan-predictive-text/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.