MochiCore
Adapter training that fits on one GPU
The trainer the rest of the lab runs on. Adapter families, quantised paths and a config format you can read out loud.
- Category
- Training
- Status
- Stable
- Language
- Python
- Licence
- MIT
- Stars
- 3.6k
- Since
- 2026
MochiCore is the oldest thing in the lab and the reason the lab exists. It began as a script that made one fine-tune reproducible and turned into the trainer everything else here depends on.
Design rules
- A config is a document. If you cannot read a run config and predict what it will do, the config format has failed.
- Resume is not optional. Consumer GPUs get interrupted. A trainer that cannot resume exactly is a trainer that loses a weekend.
- Adapters are pluggable. Adding a new family should touch one file.
mochi train configs/tagger-v1.toml --resume auto
Status
Stable. New work happens in adapters and quantised paths, not in the core loop.
References
- [1]LoRA: Low-Rank Adaptation of Large Language Models. Hu et al., ICLR 2022.
Published artefacts
Writing
Hello world
The first post. Mostly here so the blog has something to render while the real writing gets started.
Related projects
BluzenTagger
Multi-label tagging for illustration
A tagger over a 12k-tag vocabulary that holds up on the long tail, where general-purpose vision models quietly collapse into the fifty most common labels.
- Language
- Python
- Stars
- 1.4k
- Licence
- Apache-2.0
- Category
- Research
PrismStream
Flow-matching generation with region control
A text-to-image stack built around region conditioning, so a prompt can say where something belongs rather than only what it is.
- Language
- Python
- Stars
- 2.1k
- Licence
- Apache-2.0
- Category
- Research
Aurora Corpus
The dataset everything else is trained on
An 18M-image illustration corpus with tags, captions and quality bands, plus the whole filtering pipeline that produced it.
- Language
- Python
- Stars
- 980
- Licence
- CC-BY-4.0
- Category
- Data