Skip to content

SprocketLab/compositional-something

Repository files navigation

Experiment Layout

Repository modules are organized into:

  • core: shared addition/composition pipeline utilities
  • self: self-improvement experiment entrypoints
  • meta: meta self-improvement experiment entrypoints
  • launchers: Slurm/shell launch scripts
  • legacy/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 The pre-refactor standalone addition monolith was removed after checkpoint tag pre-outdated-cleanup-20260618-204614.

Canonical Commands

Run self-improvement:

python -m self.self_improvement --help

Run self-improvement composition wrapper:

python -m self.self_improvement_composition_error_experiment --help

Run multiplication self-improvement:

python -m self.multiplication_self_improvement --help

Run run-length self-improvement:

python -m self.run_length_self_improvement --help

Plot 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 --help

Artifact 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.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Packages

 
 
 

Contributors