-
Notifications
You must be signed in to change notification settings - Fork 0
Home
Speculate is a marimo notebook suite that accelerates parameter inference on disc-wind outflow spectra computed with Sirocco. It pairs a Gaussian-process spectral emulator with a full MLE-and-MCMC inference workflow, plus a lightweight neural-network alternative for quick-look fits.
If this is your first time using Speculate, work through these pages in order:
- Installation — Local environment setup with Conda or pip.
- Quick Start — The shortest path from a working install to a fitted spectrum.
- Home Tab — The landing page that exposes the rest of the suite.
The shared Troubleshooting, FAQ, and Glossary page collects cross-cutting notes that do not belong to a single tool.
Speculate is designed to run both locally and as a public HuggingFace Space. The two deployments expose different sets of tools, deliberately:
| Mode | Tools exposed | Intended use |
|---|---|---|
| Local install | Home, Model Downloader, Grid Inspector, Training Tool, Inference Tool, Quick Fit, Benchmark Suite | Full workflow: downloads, GP training, MCMC inference, and benchmarking |
| HuggingFace Space | Home, Grid Inspector, Quick Fit | Lightweight browsing of grids and pretrained quick-look fitting |
The hosted Space cannot train new emulators, run MCMC, or execute the benchmark suite. Use a local install if you need any of those workflows.
- Home Tab — Landing page and tool navigation.
- Model Downloader — Download Sirocco spectral grids and pre-trained emulator models from HuggingFace.
- Grid Inspector — Browse raw Sirocco spectra and overlay observations.
- Training Tool — Train PCA + Gaussian-process emulators.
- Inference Tool — Fit a trained emulator to a spectrum with MLE and MCMC.
- Quick Fit — Train or load a lightweight surrogate, then fit by χ².
- Benchmark Suite — Tier 1 / 2 / 3 emulator quality reports.
- Developer Notes — Contributing conventions and wiki maintenance.
🚨 Disclaimer: Most documentation has been curated using high-level reasoning LLMs (GPT5.4xHigh/GPT5.5xHigh) analysing the codebase from a human-curated plan prompt. The outputs were verified/audited by other high-level LLMs acting as judges (Claude Opus 4.7 medium/Claude Sonnet 4.6 High). The documentation should therefore, to a high degree, reflect Speculate's operations, but be aware that LLMs can make mistakes!
Speculate Tabs
Development