Quickly compare responses from different LLM models side-by-side. Use when exploring model behavior, comparing finetuned models to base models, testing prompts across models, or investigating how different system prompts affect responses.
Create clear matplotlib figures following project conventions. Use for visualizing experiment results with grouped or stacked bar charts. Saves figures to a figs/ subdirectory.
Generate interactive Streamlit apps to browse and visualize experimental results from CSV or parquet files. Use when users want to quickly explore data row by row.
Build polished SvelteKit data-viewer apps. Covers design system, component patterns, theming, charts, and human preferences learned from iterative feedback on the explanation-viewer.
Fine-tune LLMs using the Tinker API. Covers supervised fine-tuning, reinforcement learning, LoRA training, vision-language models, and both high-level Cookbook patterns and low-level API usage.