Community Grant Request: reframe — Cognitive Reframing Assistant

Summary

reframe is an open-source, privacy‑first Gradio Space that helps people practice cognitive reframing. It detects common thinking patterns (e.g., all‑or‑nothing, catastrophizing) and offers brief, supportive alternatives. The Space is educational self‑help, not therapy.

Why Reframing Matters (AvPD)

Cognitive reframing is a practical, evidence‑informed skill for reducing distress and avoidance. For people who live with strong social threat sensitivity (including many with Avoidant Personality traits/AvPD), low‑friction practice can soften self‑criticism, counter catastrophizing, and encourage graded approach behaviors. A free, judgment‑free tool lowers barriers to trying small cognitive shifts between sessions or when care is unavailable.

What the Space Provides

• Conversational assistant with streaming, gentle tone, and short, actionable guidance
• Analysis panel: likely cognitive patterns, concise reframe, and similar examples
• Learn tab: readable primers for 13+ cognitive distortions (EN/ES)
• Login with HF (OAuth) + per‑user daily caps to keep costs predictable
• Privacy by design: no message content stored; only pseudonymous, aggregated counts

Resources Requested

• GPU (L4/T4/A10G equivalent): to run open‑weight instruct models with low latency and avoid external API costs.
• Persistent storage (small): /data for privacy‑preserving counters/metrics and model cache (HF_HOME=/data/.huggingface) to speed restarts.

Technical Plan

• Gradio 5 app (app.py) with pinned deps; queue enabled for Spaces/ZeroGPU
• Agent built on huggingface_hub.InferenceClient with token streaming
• Deterministic local heuristics for distortion detection; templated examples
• Owner‑only controls (model dropdown, limits); OAuth‑gated visibility

Safety, Privacy, and Ethics

• Clear disclaimers: self‑help education, not medical advice; crisis resources shown
• No PII storage; only aggregated daily counters (paths configurable to /data)
• Open code and prompts; community feedback welcomed in the public grant thread

Milestones (90 days)

1. Enable on‑Space open‑weights pathway (GPU) with vetted instruct models and evals
2. Add AvPD‑oriented micro‑skills (graded exposure prompts; self‑compassion cues)
3. Expand locales (ES/PT/FR) + accessibility pass (keyboard/contrast)
4. Release evaluation dashboard (usage, dropout, helpfulness, safe‑completion rate)

Success Metrics

• Weekly active users, median sessions per user, 7‑day retention
• “Helpful” rating ≥ 4/5 on ≥ 60% of sessions (voluntary)
• Reduction in detected distortion intensity across a session (heuristic proxy)

Sustainability and Community

The project is MIT‑licensed and maintained in a public repo. Grant resources make it free to use and easy to duplicate/fork. We welcome contributions from clinicians, researchers, and the HF community. This request is public and we are happy to discuss scope, safety, and evaluation openly.