README.md

---
title: AnimoFlow Rewriter Probe
emoji: 🔤
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
short_description: Multilingual prompt rewriter probe (Qwen + RAFSL)
hardware: zero-a10g
suggested_hardware: zero-a10g
preload_from_hub:
  - "Qwen/Qwen2.5-1.5B-Instruct"
  - "google/gemma-3-1b-it"
  - "sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2"
---

# AnimoFlow Rewriter Probe

Phase-0 verification of the multilingual prompt-rewriter front-end planned for AnimoFlow's text-to-motion pipeline. Loads a small instruction-tuned LLM + a multilingual MiniLM retriever over the HumanML3D corpus and rewrites arbitrary user input into HumanML3D-style English captions.

- **Model:** `Qwen/Qwen2.5-1.5B-Instruct` by default (override via `MODEL_REPO` env / Space secret)
- **Retriever:** `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2`
- **Corpus:** HumanML3D captions (52K unique)
- **Hardware:** ZeroGPU (A10G burst, PRO-account free tier)

## Usage

- **Gradio UI** — type a motion prompt in any language, press *Rewrite*.
- **API** — `gradio_client.Client("AnimoFlow/rewriter-probe").predict(prompt, api_name="/rewrite")` returns `{rewritten, examples, latency_s, ...}`.

## Purpose

This Space is a **measurement probe** for the AnimoFlow rewriter plan, not a production service. Plan: `vault/wiki/concepts/Prompt rewriter front-end plan` in the AnimoFlow vault.