An interactive educational tool that teaches LLM post-training techniques through a hands-on storage I/O workload classification task. Built for the SNIA DSN Post-Training Webinar.
The guided tour walks through a progression of techniques, each building on the last:
- Prompt Engineering -- basic prompts, few-shot learning, and their limitations
- RAG -- retrieval-augmented generation with vector search
- SFT -- supervised fine-tuning with LoRA adapters
- DPO -- direct preference optimization
- GRPO -- group relative policy optimization (reinforcement learning)
Each stop includes interactive demos, visualizations, and "under the covers" deep dives into how the techniques work.
app/ React frontend (Vite + Tailwind + Zustand + D3)
scripts/ Python training pipeline (SFT, DPO, GRPO)
notebooks/ Jupyter notebooks for Colab
cd app
npm install
npm run devThe app runs entirely in the browser with precomputed data -- no backend or GPU required.
To train your own models, see scripts/README.md for the full pipeline:
- Generate synthetic dataset
- SFT with LoRA
- DPO alignment
- GRPO reinforcement learning
- Export artifacts
- Convert to ONNX for browser inference
All scripts run on a free Colab T4 GPU. See WALKTHROUGH.md for detailed instructions.
| Notebook | Purpose |
|---|---|
| Post_Training_Pipeline.ipynb | Full SFT, DPO, GRPO pipeline |
| Traditional_ML_Comparison.ipynb | XGBoost/RF baseline comparison |
| Realistic_LLM_Use_Case.ipynb | Unstructured error log classification |
The four-quadrant optimization framework used throughout this app is adapted from "A Survey of Techniques for Maximizing LLM Performance" by Colin Jarvis and John Allard, presented at OpenAI DevDay (November 2023). Also available in the OpenAI documentation.