Own the training path
Train a tokenizer and a transformer from your own data, beginning with random weights. The toolkit supports single-GPU and distributed training, plus a desktop studio for exploring the process.
03 / Model research
locallm
TRAIN · MEASURE · REPEAT
From random weights
to a model of your own.
A transformer training toolkit covering tokenization, training, exact checkpoint resumption, and inference, with an inspectable experimental record.
View repository92.9Mparameters in the recorded model
3.70BEnglish pretraining tokens
2.180final validation loss in that run
From the public project record reviewed 08 October 2026. View source and measurement context ↗
Inside the project
Built around
the details.
Resume the actual experiment
Checkpoints restore optimizer and random-number state. Resumption checks reject changes to the schedule, tokenizer, or dataset so a resumed run does not silently become a different experiment.
Keep the negative results
Predictions are registered before experiments, and the findings record whether they held. The public history includes failed transfer experiments and a headline withdrawn after a data-contamination audit.
What the results mean
Training a model is not the same as demonstrating reliable reasoning. A previously reported tie with Phi-4-mini was withdrawn after contamination was found. The model’s clean program-synthesis results were much weaker. The linked research record retains the correction and the unsuccessful experiments.
Follow the evidence