TASK.md

PaperProf — Task List

Build Small Hackathon — June 5–15, 2026

✅ Done

• Project structure created (app.py, core/, model/, ui/)
• core/parser.py: PDF text extraction with PyMuPDF
• core/chunker.py: paragraph-based text chunking with min/max word caps
• core/questioner.py: LLM-based question generation with professor prompt
• core/evaluator.py: LLM-based answer evaluation with tutor prompt
• model/llm.py: MiniCPM4-8B singleton wrapper via HuggingFace Transformers (bfloat16, chat template)
• app.py: full Gradio UI with PDF upload, question generation, answer evaluation
• requirements.txt: all dependencies listed
• README.md: project description and file structure
• .gitignore: venv, models, cache excluded
• ZeroGPU enabled on HuggingFace Space (free RTX Pro 6000 Blackwell)
• Space live at: https://huggingface.co/spaces/build-small-hackathon/PaperProf

🔲 To Do

🧪 Testing & Bug Fixes

• End-to-end test: upload real PDF → load → generate question → answer → get feedback
• Handle edge cases: empty PDF, scanned PDF (no text), very short PDF
• Handle model loading errors gracefully in the UI
• Test with multiple PDF types (slides, textbook chapters, lecture notes)
• Fix any ZeroGPU cold start issues (model takes time to load)

🎨 UI/Design (Badge: Off-Brand)

• Replace default Gradio theme with fully custom CSS using gr.Server
• Add PaperProf logo and branding
• Add progress indicator when model is generating
• Add session score tracker (X/Y correct answers)
• Add difficulty selector (Easy / Normal / Hard mode)
• Add language selector (French / English)
• Show source chunk used for the question (collapsible)
• Add end-of-session summary screen with score and weak areas
• Make the UI mobile-friendly

🧠 ML/Model Improvements

• Add quiz modes: Quick (5 questions) / Full session / Brutal mode
• Implement adaptive difficulty: revisit failed concepts, skip mastered ones
• Add chunk relevance scoring to pick the most important chunks first
• Support multiple PDF uploads in one session
• Add support for plain text (.txt) and markdown (.md) files
• Cache parsed chunks in session state to avoid re-parsing on reload

🏅 Bonus Quests

• 🔌 Off the Grid: verify zero external API calls (all local/ZeroGPU)
• 🎯 Well-Tuned: fine-tune MiniCPM on educational Q&A data using Modal credits ($250)
  • Find/create educational Q&A dataset on HuggingFace
  • Write fine-tuning script with LoRA/QLoRA
  • Run fine-tuning on Modal GPU
  • Publish fine-tuned model on HuggingFace under build-small-hackathon org
  • Update model/llm.py to use fine-tuned model
• 🎨 Off-Brand: fully custom Gradio frontend (see UI section above)
• 🦙 Llama Champion: switch inference to llama.cpp runtime
  • Download GGUF version of MiniCPM4-8B
  • Replace transformers pipeline with llama-cpp-python
  • Test performance vs transformers
• 📡 Sharing is Caring: export and share agent trace on HuggingFace Hub
• 📓 Field Notes: write blog post about what we built and learned (BLOG.md)
  • Document architecture decisions
  • Include benchmark results (speed, quality)
  • Publish on HuggingFace blog or personal blog

📦 Submission Checklist (deadline: June 15, 2026)

• App running and stable on HuggingFace Space
• Demo video (~2 minutes) showing full flow: upload → question → answer → feedback
• Social media post (LinkedIn + Twitter) with Space link and demo
• Blog post / Field Notes published (BLOG.md)
• Submission form filled on HuggingFace
• All badge requirements verified