dramatis v4 runs (preliminary)
Training runs of the v4 experiments of dramatis, a small research project on small GPT-2-style
language models (0.8M–100M parameters) trained on controlled corpora of generated children's
stories, looking for signatures of emergent misalignment: how a model builds a "persona" for a
class of character names, and how a narrow finetune spreads through it. The experiment plan is
docs/2026-09-22_EXPERIMENT_PLAN.md in the project repository (github.com/kjsatz/dramatis).
Preliminary. Runs appear here as they finish on the training machines; results are not yet analysed, checked or published. A paper and a curated release are forthcoming.
Contents
runs/<run id>/: one directory per pretraining or finetune run:config.json(with the git commit, seed and data recipe hashes), the training log and summary, the in-training probe results (probes.jsonl), and the checkpoints, as weights only (the optimizer state is not retained; every run can be recreated from its config, commit and seed).box-<k>/logs/v4/: the driver's logs and the evaluation outputs of training machine k.
Checkpoints load with dramatis.train.model_from_checkpoint from the project repository.
Where the files are
The repository takes at most 20,000 files per branch, so the four training machines' files are
split across branches: machines 1 and 4 on main, machines 2 and 3 on the branches box-2 and
box-3 (each branch holds only that machine's files; main also holds about 2,250 of machine 2's
files from before the limit was reached). For everything, download main and then each branch into
the same directory, e.g. with the project's scripts/hf_runs.py pull --revision box-2.
Training data and terms
The training stories were generated by Google's Gemma model google/gemma-4-26B-A4B-it. The
models here were trained on that generated text and may therefore be Model Derivatives of Gemma
within the meaning of the Gemma Terms of Use. They are provided under, and their use is subject
to, the Gemma Terms of Use and the
Gemma Prohibited Use Policy. Gemma is
provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms.