datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.
Paper ·
DOI ·
PDF ·
Code
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… See the full description on the dataset page: https://huggingface.co/datasets/lucky-verma/grokking-diagnostics-runs.icml17708-grok-grokking-repro
To Grok Grokking — independent reproduction
This package independently reproduces and audits the five requested claims of
To Grok Grokking: Provable Grokking in Ridge Regression by Mingyue Xu, Gal
Vardi, and Itay Safran (ICML 2026 paper 17708; OpenReview 5nNNVY8NW4; arXiv
2601.19791, v3). Reproduction date: 2026-07-16.
Outcome first
The ridge mechanism and its hyperparameter predictions reproduce strongly, but
the package does not support every claim without… See the full description on the dataset page: https://huggingface.co/datasets/YMRohit/icml17708-grok-grokking-repro.lens-loss-grokking-experiment
lens-loss-grokking-experiment — checkpoints
Final model checkpoints for the experiments in
brendanlong/lens-loss-grokking-experiment
(deep supervision vs grokking: LN-scoped chronic instability, weight-decay
circuit pruning, and failure isolation). Training curves:
public wandb project.
Layout: grok_lens/<run_name>/final.pt (modular-arithmetic runs, ~1.7 MB
each, with .json metadata sidecars) and lego/<run_name>/step_*.pt
(S3 multi-hop composition runs). Run names encode the… See the full description on the dataset page: https://huggingface.co/datasets/brendanlong/lens-loss-grokking-experiment.repro-evidence-grokking-ca-local-rules
Evidence trail — Grokking phase transitions in learning local rules with gradient descent
Full evidence for an automated claim-by-claim audit of
Grokking phase transitions in learning local rules with gradient descent, produced by
Lemma, an AI-scientist pipeline built
for re:AGENT (Founders Inc, Aug 15–16 2026).
Verdict: 5 supported / 0 falsified / 1 inconclusive
of 6 extracted claims. Judge verdict: PASS (5/5).
Claim
Title
Verdict
C1
Critical exponent in 1D… See the full description on the dataset page: https://huggingface.co/datasets/Papajams/repro-evidence-grokking-ca-local-rules.icml17708-grok-grokking-reprorepro-grokking-ridge-5nNNVY8NW4-bundlerepro-grokking-ridge
Repro — To Grok Grokking (5nNNVY8NW4): ridge-regression grokking
ICML 2026 Agent Reproduction Challenge — Claim-Closure Agent 02.
Paper: To Grok Grokking: Provable Grokking in Ridge Regression (OpenReview 5nNNVY8NW4, arXiv 2601.19791).
An independent first-principles finite-d study of whether end-to-end
grokking (overfit → delayed poor generalization → eventual low error) occurs
in over-parameterized realizable ridge regression under GD + constant weight
decay. Every number is… See the full description on the dataset page: https://huggingface.co/datasets/pranaysuyash/repro-grokking-ridge.to-grok-grokking-reproduction-code
To Grok Grokking — independent reproduction code
Self-contained scripts for an independent reproduction of ICML 2026 paper
#17708, To Grok Grokking: Provable Grokking in Ridge Regression.
theory_audit.py: bounded-Rademacher finite-dimensional audit of Theorems
4.1, 4.2, and 4.4–4.6, including condition-relaxation controls.
ridge_gpu_sweep.py: paper-scale spectral GPU reproduction of the Figure 2
weight-decay and sample-size panels.
relu_gpu_sweep.py: declared-Gaussian… See the full description on the dataset page: https://huggingface.co/datasets/visv-Bro/to-grok-grokking-reproduction-code.openinterp-39-grokking-retrospective
nb39 — Grokking retrospective on nb37 DPO checkpoints
Tests whether DPO on Qwen3.6-27B (nb37) shows phase-transition learning detectable via probes.
Hypothesis
Probe AUROC for preference-shifted output undergoes phase transition during DPO training, before greedy decoding diverges.
Result
See FINAL_VERDICT.json. Grokking signal: {verdict.get("grokking_signal", "undetermined")}.
Phase transition ratio: {verdict.get("phase_transition_ratio", "N/A"):.2f}.… See the full description on the dataset page: https://huggingface.co/datasets/caiovicentino1/openinterp-39-grokking-retrospective.grokking-checkpointsopeninterp-41v2-grokking-extended
nb41 v2 — Grokking forward-only on extended DPO checkpoints
Resolves nb41 v1 ambiguity (ratio=1.74) using nb37 v2 extended training (10 checkpoints across 200 steps with -0.23 loss descent vs v1's 4 checkpoints across 80 steps with -0.04 descent).
Methodology: forward-only on (prompt + chosen), capture L31/L55 at end-of-think, score with FG+RG probes, fresh-probe AUROC progression.
Key fix: strip .language_model. from saved LoRA keys before PeftModel.from_pretrained() (Qwen3.6… See the full description on the dataset page: https://huggingface.co/datasets/caiovicentino1/openinterp-41v2-grokking-extended.openinterp-41-grokking-forward-only
nb41 — Grokking forward-only (nb39 v2 with Qwen3.6 LoRA key fix)
Fixes the bug discovered in nb40: Qwen3.6-27B PEFT save creates keys with .language_model. infix; PeftModel.from_pretrained against dense reload silently fails (zero LoRA effect).
Uses forward-only methodology: feed prompt + chosen from nb37 pairs.json through each checkpoint, capture L31/L55 at end-of-think, score with FabricationGuard + ReasonGuard probes. ~10 min compute.
See FINAL_VERDICT.json for results.
Grokking_checkpoints_antigrokp2-etf-grokking-detection-monitor-resultsgrok-grokking-ridge-repro-codeGPT2_GROKKING