BurnyCoder/grokking-modular-addition-site-data

Grokking modular addition — website data

The data read by the website https://burnycoder.github.io/visualizing-grokking-modular-addition/ (code: https://github.com/BurnyCoder/visualizing-grokking-modular-addition): per-checkpoint measurements, loss curves, Fourier spectra, activation geometry and float16 weights of the mainline run of Nanda et al. 2023, Progress measures for grokking via mechanistic interpretability (arXiv:2301.05217), and of our 22 retrained runs (their full checkpoints are in https://huggingface.co/BurnyCoder/grokking-modular-addition-runs). Written by the export_site_data phase; docs/data-formats.md in the code repository documents every file.

data/ contains:

• manifest.json: every run (key frequencies, phases, checkpoint steps) and every array (file, dtype, shape, sha256);
• runs/<id>/<name>.bin: little-endian arrays (float32, float16, int16 or quantised uint8), one row per checkpoint, or per logged step for the loss curves and update norms;
• runs/<id>/w_<step>.bin: float16 weights at selected checkpoints, for the in-browser model;
• train_mask_seed<s>.bin: the train/test split of each seed;
• summary.json: the numbers and tables quoted in the website's prose.

Licence and attribution:

• data/runs/paper/w_*.bin are float16 copies of the weights of Nanda et al. 2023 (arXiv:2301.05217), downloaded from https://huggingface.co/callummcdougall/grokking_full_run_data at revision 80e5dbb5744dd3e958e45e56f7562a4d50c7400c. No licence was found for them; they are re-published with attribution so the website can run the paper's network in the browser.
• Everything else (the arrays we computed, and our runs' weights) is MIT-licensed.

Code, website and reports: https://github.com/BurnyCoder/visualizing-grokking-modular-addition. Author: Libor Burian.