---

title: InnerSpace emoji: 🧠 colorFrom: purple colorTo: gray sdk: gradio sdk_version: 6.17.3 app_file: app.py python_version: "3.12" short_description: Local-first cognitive journal & reflection coach pinned: false tags:

• build-small-hackathon
• backyard-ai
• openbmb
• openai
• well-tuned
• off-brand
• off-the-grid
• tiny-titan
• mental-health
• journaling
• modal

---

Inner 🧠 Space

InnerSpace is a private, local-first cognitive journal and AI reflection companion. It runs a fine-tuned 1.2B parameter language model inside the Hugging Face Space runtime. There is no serverless inference fallback, so journal text is not sent to an external inference API.

The model analyzes journal entries through the lens of Cognitive Behavioral Therapy (CBT): surfacing emotions, identifying affected life areas, flagging cognitive distortions, and responding with a gentle reflective question to help the writer think more clearly.

InnerSpace is a reflective journaling tool, not medical advice, diagnosis, crisis counseling, or a replacement for a licensed mental-health professional. If someone may be in immediate danger or crisis, they should contact local emergency services or a crisis hotline.

Links

• Demo: Video
• Social Post: LinkedIn
• Github Repo: awilliams88/innerspace
• Hugging Space: build-small-hackathon/innerspace
• Fine-tuned Model: build-small-hackathon/inner-space-1b-sft-cbt

Hackathon Alignment

What It Does

Write or upload a journal entry (.txt or .md) and set your current distress level. InnerSpace will return a structured reflection in six parts:

---

Fine-Tuned Model

The inference engine is powered by a QLoRA-adapted version of openbmb/MiniCPM5-1B-SFT, trained specifically on CBT reflection patterns.

Why fine-tune instead of prompting? The base model is general-purpose. Fine-tuning teaches it the core CBT output structure and vocabulary — producing more consistent, therapeutically-grounded responses without relying on long system prompts. The current app extends that flow with a balanced reframe, a tiny next step, and distress-level context.

Training details:

• Method: QLoRA (4-bit NF4 quantization + LoRA adapters on attention layers)
• Hardware: NVIDIA A10G GPU via Modal.com
• Dataset: 17 structured CBT journal entries plus 19 multi-turn follow-up coaching examples (including off-topic deflection and redirection)
• Output format: six sections aligned with the app UI — emotions, life areas, cognitive distortions, balanced reframe, tiny next step, and reflection
• Follow-up behavior: brief second-turn coaching for self-critical replies without hidden reasoning tags or business-style metrics
• Steps: 220 with a rank-16 LoRA adapter and 1536-token examples

The fine-tuned LoRA adapter is published at build-small-hackathon/inner-space-1b-sft-cbt and is loaded automatically on top of the base model at Space startup.

---

Inference Architecture

User Input (text or file)
        │
        ▼
┌─────────────────────┐
│    Gradio UI        │  ui.py — dark-violet mindful dashboard
└──────────┬──────────┘
           │
           ▼
┌─────────────────────┐
│  Analyzer           │  analyzer.py — prompt construction & ZeroGPU dispatch
└──────────┬──────────┘
           │
     ┌─────┴──────┐
     ▼            ▼
┌─────────┐  ┌──────────┐
│Inference│  │  Parser  │  inference.py — model execution
│ Engine  │  │  Engine  │  parser.py — file reading & section splitting
└────┬────┘  └──────────┘
     │
     └── ZeroGPU / local runtime: base model + LoRA adapter via PeftModel

Inference priority:

1. ZeroGPU — loads MiniCPM5-1B-SFT in bfloat16 and applies the fine-tuned LoRA adapter via PeftModel. Runs on an NVIDIA A10G in the Space.
2. Privacy-first failure policy — if local inference fails, the app returns a clear error instead of routing journal text to a serverless API.
3. Error — if local execution fails, the UI returns a clear error message. No silent failures.

---

Local Development

Setup:

./run.sh setup

Run locally:

./run.sh app

This launches through app.py so Gradio receives the custom theme and CSS.

Quality checks (Ruff formatting, Ruff linting, Pyright type checking, Python compilation):

./run.sh verify

---

Codebase

Root

env/ — App infrastructure

core/ — Business logic

ui/ — Presentation

modal/ — Remote fine-tuning

Project files

---

Tech Stack

• Model: openbmb/MiniCPM5-1B-SFT + custom LoRA adapter (build-small-hackathon/inner-space-1b-sft-cbt)
• Fine-tuning: QLoRA via peft + trl SFTTrainer on Modal A10G
• Inference: transformers + peft (PeftModel) + accelerate
• UI: Gradio 6 with custom CSS
• Hosting: Hugging Face Spaces (ZeroGPU)
• Sponsor: OpenBMB — MiniCPM model family

---