Repository modules are organized into:
core: shared addition/composition pipeline utilitiesself: self-improvement experiment entrypointsmeta: meta self-improvement experiment entrypointslaunchers: Slurm/shell launch scriptslegacy/weak_to_strong: archived weak-to-strong launch scripts and wrappers around the canonical addition pipeline. The pre-refactor standalone addition monolith was removed after checkpoint tagpre-outdated-cleanup-20260618-204614.
Run self-improvement:
python -m self.self_improvement --helpRun self-improvement composition wrapper:
python -m self.self_improvement_composition_error_experiment --helpRun multiplication self-improvement:
python -m self.multiplication_self_improvement --helpRun run-length self-improvement:
python -m self.run_length_self_improvement --helpPlot a self-improvement figure from one or more completed runs:
python -m self.plot_self_improvement_figure RUN_DIR [RUN_DIR ...]Run meta self-improvement:
python -m meta.train_meta_self_improvement_rope --helpArtifact outputs default to artifacts/.
The workshop defaults now target HuggingFaceTB/SmolLM2-360M and use bf16
automatically on CUDA when neither --bf16 nor --fp16 is specified.
The Slurm launchers prefer a local artifacts/models/SmolLM2-360M snapshot
when it exists and otherwise fall back to the Hugging Face model id
HuggingFaceTB/SmolLM2-360M.