lucky-verma/grokking-diagnostics-runs

Grokking Diagnostics Runs

Per-run training records and aggregate fits backing:

Weight Decay Regimes in Grokking Transformers: Cheap Online Diagnostics Lucky Verma. Independent Researcher. 2026.

Companion code and paper PDF: https://github.com/lucky-verma/grokking-diagnostics

Contents

The paper provenance indexes 1,792 paper-run records: 1,442 records from the main paper-integrated run tree plus 350 cross-architecture scope-probe records. This dataset repository also includes convenience subset mirrors, so the repository contains more JSON files than unique paper-run records.

Subset	Path	JSON files	Description
Main paper-run mirror	transformer/	1,442	Paper-integrated modular-arithmetic runs, including the MLP scope-probe records used in the manuscript provenance.
LSTM scope probe	cross-arch/lstm/	70	E14 4L LSTM h=512 mod_add probe, 7 weight-decay values x 10 seeds at 10K epochs.
Mamba scope probes	cross-arch/mamba/	210	E15 Mamba grids: expand-4 mod_add, expand-2 mod_add, and expand-4 mod_mul.
Canonical convenience view	canonical/	50	Canonical 4L8H d=128 mod_add lambda=1.0 runs used for the two-phase trajectory.
Multi-task convenience view	e9-multitask-transformer/	280	Horizon-matched four-operation replication.
Causal-intervention view	e12-interventions/	60	Head re-initialization and weight-clipping paired intervention records.
Long-horizon checkpoint views	e13-crossseed-checkpoints/, e13-canonical-checkpoints/	5	Cross-seed and canonical long-horizon checkpoint traces.
Aggregate fits and provenance	aggregates/	10 JSON + 1 markdown	Rich aggregate fits, claim-to-data map, and coverage dashboard.

Per-run JSON schema

Each per-run JSON contains a configuration block, a training-history list, final train/test metrics, elapsed time, parameter count, grokking epoch if present, and status metadata. Cross-architecture probes include architecture-specific configuration fields such as arch, n_layers, d_model, expand, and d_state where applicable.

Example fields:

{
  "config": {
    "p": 97,
    "wd": 0.05,
    "lr": 0.001,
    "epochs": 10000,
    "batch_size": 512,
    "seed": 42,
    "task": "mod_add",
    "arch": "Mamba",
    "label": "..."
  },
  "history": [
    {
      "epoch": 10000,
      "train_acc": 1.0,
      "test_acc": 0.99,
      "train_loss": 0.0,
      "test_loss": 0.0,
      "weight_norm": 0.0
    }
  ],
  "grok_epoch": 3000,
  "final_train_acc": 1.0,
  "final_test_acc": 0.99,
  "elapsed_sec": 0.0,
  "n_params": 957952,
  "status": "done"
}

Transformer records include attention-derived diagnostics where available: mean pairwise head cosine similarity, entropy standard deviation across heads, and related logged quantities.

Provenance

Every figure and table cell in the paper traces to an aggregate JSON through aggregates/paper_sources.json. The coverage dashboard aggregates/COVERAGE.{md,json} reports:

verified: 57
mismatch: 0
pending: 0
missing: 0
unknown: 0

Quick start

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="lucky-verma/grokking-diagnostics-runs",
    repo_type="dataset",
    local_dir="./data/raw_jsons",
)

For code, verification scripts, figures, and Lean 4 proofs, use the companion repository:

git clone https://github.com/lucky-verma/grokking-diagnostics.git
cd grokking-diagnostics
python scripts/download_dataset.py
python scripts/verify_numerical_claims.py

Citation

@article{verma2026grokking,
  title  = {Weight Decay Regimes in Grokking Transformers: Cheap Online Diagnostics},
  author = {Verma, Lucky},
  year   = {2026}
}

Use the companion repository's CITATION.cff for machine-readable citation metadata. The citation can be updated with a public preprint identifier or DOI after one is assigned.

License

Dataset: CC-BY-4.0. Companion code: Apache-2.0.

Contact

Open an issue at https://github.com/lucky-verma/grokking-diagnostics/issues or email luckyv1@umbc.edu.