Hugging Face's logo

Spaces:
AndaiMD
/
brainbench
Sleeping

App Files Community
brainbench / README.md
AndaiMD's picture
AndaiMD
README
5dfbe24
preview code
|
raw
history blame contribute delete
2.29 kB
metadata
title: Brainbench
emoji: 🦀
colorFrom: gray
colorTo: purple
sdk: docker
pinned: false
short_description: Implementing the braingpt model

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

This app currently uses GPT-2 as the model engine for generating and evaluating abstracts.

🚀 Features

  • 🧪 Presents users with neuroscience abstracts (either original or altered).
  • ✅ Users decide if an abstract is AI-modified and rate their confidence.
  • 🤖 GPT-2 model evaluates the same abstract.
  • 📊 Results compare user guesses vs model output in a clear, styled interface.
  • 📂 Backed by a curated dataset hosted on Hugging Face Datasets.
  • 🛠️ Fully Dockerized FastAPI application deployed via Hugging Face Spaces.

📦 Tech Stack

  • Frontend: Jinja2 + Tailwind-style CSS
  • Backend: FastAPI
  • Model: gpt2 via Hugging Face Transformers
  • Hosting: Hugging Face Spaces (Docker SDK)
  • Dataset: Custom neuroscience benchmark (Parquet format)

🔧 Endpoints

  • /: Home page
  • /start: Begin a new session (random abstract trials)
  • /trial: Active abstract assessment
  • /submit-trial: Submit a response, compare with model
  • /results: View summary of performance
  • /predict: POST endpoint used internally to get GPT-2 output

🧠 Example Use Case

  1. A neuroscience researcher lands on the app.
  2. They read an abstract and guess whether it’s been altered.
  3. They rate their confidence.
  4. GPT-2 also analyzes the abstract.
  5. After 3 rounds, the app displays a comparison: who got what right?

📌 Current Model

  • Model: gpt2
  • Reason: Fast, lightweight, deployable in CPU-only Spaces
  • Coming Soon: BrainGPT (a PEFT fine-tuned model on neuroscience texts)

🗺️ Roadmap

  • Replace GPT-2 with BrainGPT once optimized
  • Add leaderboard tracking
  • Improve GPT model alignment with truth labels
  • Expand dataset with more domains

🧪 Try It Live

👉 BrainBench on Hugging Face Spaces

🙌 Credits

Built by Dr. David Andai — a general practitioner & data scientist passionate about mental health and AI.