Irutingabo/capstone_project

Umubyeyi — fine-tuned generator inference API

This Space hosts Umubyeyi's own fine-tuned emotional-support generator (BLOOMZ-560m, LoRA on real ESConv data) so the deployed Vercel app can call real generation instead of falling back to retrieval — Vercel's serverless Python functions cannot bundle PyTorch/Transformers within their 500MB size limit, but this Space (with a free ZeroGPU allocation) has no such constraint.

This is not a chat UI for end users, though the page above lets you try it directly. The programmatic contract is a named Gradio API endpoint:

POST {space_url}/call/generate  {"data": ["<message>", "en", "[]"]}
-> {"event_id": "..."}
GET  {space_url}/call/generate/{event_id}   (Gradio's async SSE result stream)

The model, LoRA adapter, and prompt format here are identical to src/finetuned_generator.py in the main repository (https://github.com/IrutingaboRaissa/UMUBYEYI) — this Space exists purely to run the same weights on infrastructure without Vercel's size limit, not to run a different model.