Software
Tools for building, training, and understanding small language models — a desktop training studio, a terminal client, and a training-loss analyzer. Each one runs locally, stays close to the underlying engine, and is built for people who actually train models rather than just consume them.
Applications
The Rootcomputer toolkit
Three focused apps that cover the lifecycle of a small language model: build and train it, talk to it, and understand how its training is going.
A local desktop workspace that wraps our h2 Python training engine: project management, tokenizer building, pretraining, SFT, DPO alignment, checkpoint testing, and kernel-level runtime visibility — without burying the engine underneath it.
- Tokenizer training
- Pretraining
- Fine-tuning
- DPO
- Chat lab
A dependency-free command-line client for interactive model use, endpoint testing, optional plugin workflows, sampling control, conversation capture, and local SFT/DPO preference-data collection. No pip install required.
- Zero dependencies
- Endpoint switching
- Plugins
- Sampling control
- Feedback capture
Reads any training log, diagnoses run health, extracts your learning-rate schedule, and forecasts where the loss curve is headed — with backtested confidence bands and the milestones worth watching. Understands pretraining and SFT, and reads whatever format your trainer prints.
- Health diagnostics
- Schedule-aware forecasting
- Regime detection
- Format-agnostic parsing
- Interactive charts
How we build
Local-first, close to the engine.
Every tool here runs on your own machine and stays close to the underlying training scripts rather than hiding them. Your data, checkpoints, and logs never leave your hardware, and the source is open so you can see exactly what each app does.