| --- |
| title: REFRAME |
| emoji: 🧠 |
| colorFrom: purple |
| colorTo: indigo |
| sdk: gradio |
| app_file: app.py |
| pinned: false |
| license: gemma |
| tags: |
| - build-small-hackathon |
| - track:backyard |
| - sponsor:modal |
| - sponsor:cohere |
| - achievement:offgrid |
| - achievement:welltuned |
| - achievement:llama |
| --- |
| |
| # REFRAME — a live CBT studio |
|
|
| **Reframe your thinking.** A live cognitive restructuring studio — talk through an unhelpful thought |
| (**by voice or text**) and watch it turn into a fairer, kinder one, live. The chat is a **Gemma 4 12B** |
| model **fine-tuned on ~42,000 therapy conversations** (QLoRA on **Modal** · H100), running locally via |
| **llama.cpp**; your voice is transcribed by **Cohere Transcribe**. Small models, fully on-device — no cloud APIs. |
| A Thought Card builds the cognitive-restructuring record in real time (distortion → evidence → your |
| balanced reframe), saved to a deck with a pattern tracker. |
|
|
| **Track:** Backyard AI — a real problem (rumination / unhelpful thinking), honestly fit to the small-model constraint. |
|
|
| ## How it works |
| 1. **Speak** a thought → **Cohere Transcribe** turns it into text. |
| 2. The **fine-tuned Gemma** reflects it back and asks Socratic questions. |
| 3. A **Thought Card** tags the distortion, gathers evidence, and you write the **balanced thought** — it's yours, so it sticks. |
| 4. Cards save to your **Deck**; the **Patterns** tab shows what recurs over time. |
|
|
| ## Tech (all ≤ 32B, runs locally) |
| - **Chat:** `google/gemma-4-12B-it` (12B), **fine-tuned (QLoRA) on mental-health counseling, empathetic-dialogue & crisis-response data** → published on the Hub. |
| - **Training:** **Modal**, single **NVIDIA H100 80GB** (unsloth · peft · trl · bitsandbytes). |
| - **Inference:** merged **GGUF Q4_K_M**, run by **llama.cpp** (Ollama locally / llama-cpp-python on the Space). |
| - **Voice:** **Cohere Transcribe** (`CohereLabs/cohere-transcribe-03-2026`) — on **ZeroGPU** here, Intel Arc XPU locally. |
| - **UI:** custom Gradio theme; private by design (no cloud APIs). |
|
|
| ## Badges |
| - ✅ **Well-Tuned** — fine-tuned, published model |
| - ✅ **Llama Champion** — llama.cpp runtime |
| - ✅ **Off the Grid** — no cloud APIs; all inference local/self-hosted |
| - *(Off-Brand: verify it qualifies — needs custom UI beyond default Gradio)* |
|
|
| ## Links |
| - 🎥 **Demo video:** https://youtu.be/OjWORvLAwq8 |
| - 📣 **Social post:** https://x.com/v_r_droids/status/2066664719822282796 |
| - 🤖 **Model:** https://huggingface.co/emailvenky/gemma4-mentalhealthbuddy-v1 |
|
|
| ## Note |
| A supportive tool, not a replacement for professional care. If a message signals crisis, REFRAME surfaces helplines immediately. |
|
|
| — Created with 🍁 in Canada |
|
|
