DynCL: Bio-Inspired Structural Plasticity as a Continual Learning Substrate for Curriculum Transformers

This repository is the publication-facing reproduction package for the DynCL paper.

It contains:

• the paper PDF;
• raw per-run evaluation JSON files;
• paper-facing aggregate JSON files;
• the linear_1d rollout audit artifacts;
• curriculum and tokenization documentation;
• lightweight scripts to regenerate the released figures and summary tables without retraining.

The package is organized for artifact reproduction first. Full retraining remains possible from the main project codebase, but is not required to rebuild the reported paper figures and tables.

Paper

• article/paper.pdf

Repository layout

• article/
  • paper PDF
• evaluation/raw_json/
  • per-run evaluation JSON files
• evaluation/aggregated/
  • paper-facing aggregates used to rebuild tables and figures
• evaluation/audit/
  • targeted linear_1d rollout audit
• figures/paper/
  • released paper figure PDFs
• figures/generated/
  • regenerated figures created by scripts/make_figures.py
• configs/
  • evaluation and run manifests
• curriculum/
  • curriculum documentation
• tokenization/
  • tokenization notes and exact helper code copied for the release
• docs/
  • provenance, reproducibility checklist, and release notes

Quick reproduction

Create the environment:

cd dyncl
mamba env create -f environment.yml
mamba run -n dyncl python scripts/make_tables.py
mamba run -n dyncl python scripts/make_figures.py

Outputs:

• tables: tables/generated/
• figures: figures/generated/

What the scripts regenerate

scripts/make_tables.py regenerates:

• main_results.csv
• main_results.md
• balanced_global.csv
• balanced_global.md
• linear_1d_audit.csv
• linear_1d_audit.md

scripts/make_figures.py regenerates:

• pareto_primary_rebuilt.pdf
• pareto_primary_rebuilt.png
• pareto_middle_rebuilt.pdf
• pareto_middle_rebuilt.png
• cl_trajectory_rebuilt.pdf
• cl_trajectory_rebuilt.png

Full retraining

This repository is not optimized as a standalone training repo. The released artifacts are sufficient to reproduce the reported paper figures and tables without rerunning expensive training.

For full retraining, use the main project codebase and the manifests in:

• configs/run_configs/
• configs/evaluation/

Data

The DeepMind Mathematics dataset is not redistributed here. See:

• curriculum/dataset_note.md

Checkpoints

Checkpoint handling is documented in:

• checkpoints/README.md

Archived paper checkpoints:

• Zenodo DOI 10.5281/zenodo.20511782

Provenance and release notes

• docs/provenance.md
• docs/release_notes.md
• docs/reproducibility_checklist.md
• docs/reproduction.md

Citation

See:

• CITATION.cff

License

This repository is released under the MIT license:

• LICENSE

The upstream dataset remains subject to its own terms.