Blog

Notes from the training loop

Launch notes, workflow changes, and practical walkthroughs of turning raw data into small, evaluated local models.

All posts

Jul 1, 20264 min read

Tuned Tensor Local: the whole loop, open source, on your machine

A free, Apache-2.0 CLI that runs the full fine-tune-and-evaluate loop on your own GPU. No cloud account, no Docker, no managed database — every artifact is a file on your disk.

Read post
Jun 30, 20264 min read

Introducing model lineage

Trace continued fine-tunes from base model to parent model to child model, with lineage trees now visible across the dashboard.

Read post
Jun 24, 20264 min read

Fine-tune Qwen3-VL 2B for OCR and document extraction

Bring image and document-page inputs into a behavior-first fine-tuning workflow that turns visual documents into structured JSON.

Read post
Jun 20, 20264 min read

Fine-tune Qwen 3.5 2B for financial sentiment analysis

A public Hugging Face dataset, a strict JSON behavior spec, one Qwen3.5-2B fine-tune, local smoke tests, and a shipped model card.

Read post
Jun 14, 20263 min read

A new labeling loop for training data

A concrete finance-news labeling pass: 1,000 accepted rows, settled cost, review flow, provenance notes, and the spec bundled with the dataset.

Read post
Jun 8, 20266 min read

Fine-tune Qwen 3.5 2B for email safety triage

One Qwen 3.5 2B fine-tune on a public 10k-row dataset that triages email and flags phishing, spam, and prompt injection. Dataset, model, and spec all published.

Read post

Want to learn more?

Explore the documentation to see how behaviour specs, runs, and evaluations work under the hood.