| --- |
| title: cali |
| emoji: 🌏 |
| colorFrom: blue |
| colorTo: indigo |
| sdk: gradio |
| sdk_version: 6.30.0 |
| app_file: app.py |
| python_version: "3.12" |
| short_description: Indonesian-English translation and CALI text generation |
| license: apache-2.0 |
| models: |
| - Sandroeth/cali-id-en-translate |
| - Sandroeth/cali-0.1B |
| --- |
| |
| # cali |
|
|
| Explore two lightweight language models by Sandroeth in an English-language interface. |
|
|
| - **Translation:** translate everyday sentences between Indonesian and English. |
| - **Explore CALI:** continue a text prompt with the experimental CALI-0.1B base model. |
|
|
| ## Inference |
|
|
| Translation follows the [official model README](https://huggingface.co/Sandroeth/cali-id-en-translate/blob/main/README.md): `[id→en]\n{text}\n→` or `[en→id]\n{text}\n→`, deterministic decoding, `repetition_penalty=1.5`, `no_repeat_ngram_size=3`, `use_cache=False`, and 64 output tokens by default. The first-sentence option matches the official example; disable it to inspect the complete generated output. |
|
|
| The base model receives a plain text prompt without a chat template. Both models use their custom architecture code at the revisions pinned in `app.py`. Float32 preserves compatibility with the translation model's RoPE implementation. Transformers 5.9.0 matches the translation checkpoint configuration. |
|
|
| The combined input and output budget is 1024 tokens, with up to 256 output tokens. Translation inputs are limited to 200 words. |
|
|
| ## Hosting and API |
|
|
| This Gradio Space uses ZeroGPU. Both models are loaded once at startup, and the inference handlers use `@spaces.GPU`. CPU execution is also supported when available. Spaces manages Gradio, PyTorch, `spaces`, and `huggingface_hub`. |
|
|
| Gradio API endpoints: `/translate` and `/complete`. MCP is enabled. Examples are cached lazily after first use. Each response includes token counts and inference time. |
|
|
| ## Limitations |
|
|
| Translation works best on everyday sentences. Very short inputs, technical topics, and long passages may be unreliable. CALI-0.1B is an experimental base model trained on limited data; continuations can repeat, drift off-topic, or be inconsistent. |
|
|
| Both models are released under Apache 2.0 by Sandroeth. |
|
|
