FedGrok/fedgrok-checkpoints

Federated grokking — model checkpoints and per-client weights

36,298 PyTorch checkpoint files, 39.9 GB, from 664 runs of the v2 multi-setup campaign of a study on whether grokking survives federated averaging (FedAvg). One archive per campaign group, filed by the axis of the paper it supports; no compression (these are already-packed float tensors and gzip recovers ~8%). Access is gated with automatic approval: click request access once, then any authenticated download works.

The study trains five setups — A: quadratic MLP on mod-97 addition (GD, wd 0); A': the same with AdamW; B: one-layer transformer on mod-113 addition (AdamW); C: transformer on S5 composition; D: quadratic MLP on S5; E: MLP on MNIST — centralised and under FedAvg, and measures when each memorises and when it grokks as the number of clients K, local epochs E, participation f and data heterogeneity vary.

Layout

path	what it is
num_clients/	the K ladder for every setup, with the 87 centralised anchors (the K=1 end) inside checkpoints_aggregation.tar — 5 archives, 251 runs, 15.6 GB
local_epochs/	local work per round, E in {10, 25, 50}, and the long E=50 runs behind the fixed point — 2 archives, 42 runs, 1.9 GB
participation/	fraction of clients per round — 2 archives, 39 runs, 1.7 GB
heterogeneity/	Dirichlet label skew, the client-size control, and the structured partitions — 4 archives, 216 runs, 12.4 GB
mechanism/	the boundary cells and the setup-D internals the mechanism analysis reads — 2 archives, 105 runs, 8.3 GB
legacy/	11 partial runs from a superseded batch; not in the run table, keep out of analyses — 1 archive, 11 runs, 0.1 GB
MANIFEST.csv	one row per archive: path, group, run/file counts, bytes, sha256, axis, figure, campaign, setups, what varies
SHA256SUMS	sha256sum -c SHA256SUMS from this directory verifies every archive
runs_v2.csv	the run table: 1,685 rows, one per run, every config and result field

Folder names are the run table's column names, not figure numbers, so they stay put when the paper is renumbered; the figure column of MANIFEST.csv carries that map. Before 8 Sep 2026 every archive sat at the repository root; those paths no longer resolve.

Which archive holds which runs

A run is identified by run_id, a content hash of its full config. The join is: runs_v2.csv maps run id -> group (plus setup, K, E, f, alpha_dir, partition, weight decay, seed, t_memo, t_first_cross, ...); MANIFEST.csv maps group -> path. Runs with checkpoint_every = 0 in the table have no checkpoints and are in no archive. Every run with checkpoint_every > 0 is in exactly one archive, complete for its save schedule (verified file by file on 8 Sep 2026).

num_clients/ — number of clients K

The K ladder for every setup, with the 87 centralised anchors (the K=1 end) inside checkpoints_aggregation.tar. Sept 2026 draft: Fig 1, Fig A1, text (B K-collapse diagnosis), text (B weight-decay control).

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
num_clients/checkpoints_aggregation.tar	aggregation	A/A'/B/C/D/E	K in {2, 5, 10, 20, 50}; alpha in {0.1, 0.2, 0.3, 0.4, 0.5, 0.01, 0.02, 0.03, 0.04, 0.06, 0.08, 0.15, 0.004, 0.006, 0.008, 0.015}; wd in {0.0, 0.1, 1.0}	174	13,740	11.80 GB	b43883ce6a86d885
num_clients/checkpoints_aggregation_alpha2.tar	aggregation_alpha2	B/E	K in {2, 5, 10, 20}; alpha in {0.4, 0.5}; wd in {0.1, 1.0}	24	480	0.78 GB	da4fd7b34aeffd0c
num_clients/checkpoints_setup_k_ladder.tar	setup_k_ladder	B/C/D/E	K in {5, 10, 20, 50}; alpha in {0.3, 0.5}; wd in {0.1, 1.0}	39	780	2.43 GB	7dc33e41af36d27f
num_clients/checkpoints_b_k20_wd_control.tar	b_k20_wd_control	B	-	3	60	0.08 GB	2f3ecf6d6981a463
num_clients/checkpoints_k_collapse_budget.tar	k_collapse_budget	B	K in {20, 30, 50}; wd in {0.1, 1.0}	11	280	0.51 GB	ee20621ad5b16903

local_epochs/ — local epochs E

Local work per round, E in {10, 25, 50}, and the long E=50 runs behind the fixed point. Sept 2026 draft: Fig 2, Fig 6.

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
local_epochs/checkpoints_local_epochs.tar	local_epochs	B/C/D/E	E in {10, 25, 50}; alpha in {0.3, 0.5}; wd in {0.1, 1.0}	36	720	1.67 GB	997dfa7598b56d8d
local_epochs/checkpoints_e50_long.tar	e50_long	C/D	alpha in {0.3, 0.5}	6	120	0.19 GB	2f1405b83178a22a

participation/ — participation f

Fraction of clients per round. Sept 2026 draft: Fig 3.

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
participation/checkpoints_participation.tar	participation	A	f in {0.2, 0.4, 0.6}	9	540	0.63 GB	71138497caf8e996
participation/checkpoints_participation_setups.tar	participation_setups	A/B/C/D/E	f in {0.5, 0.25}; alpha in {0.3, 0.5}; wd in {0.0, 0.1, 1.0}	30	780	1.09 GB	6bdb6032e779a5d6

heterogeneity/ — heterogeneity

Dirichlet label skew, the client-size control, and the structured partitions. Sept 2026 draft: Fig 4a-b, Fig 4c, Fig 4d.

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
heterogeneity/checkpoints_dirichlet_band.tar	dirichlet_band	A	K in {20, 50}; alpha_dir in {0.1, 1.0, 0.01, 10.0, 1000.0}	24	1,440	2.52 GB	63e3d7b161e8c8c4
heterogeneity/checkpoints_dirichlet_setups.tar	dirichlet_setups	A/B/C/D/E	alpha_dir in {0.1, 1.0, 0.01, 10.0, 1000.0}; alpha in {0.3, 0.4, 0.5}; wd in {0.0, 0.1, 1.0}	72	1,590	2.77 GB	0b28d408709a5989
heterogeneity/checkpoints_size_control.tar	size_control	A	K in {20, 50}; alpha_dir in {0.1, 0.01}	12	720	1.42 GB	862dd84b05f2231b
heterogeneity/checkpoints_partitions.tar	partitions	A/A'/B/C/D/E	K in {5, 10, 20, 50}; partition in {coset, target, operand, dirichlet, label_block}; alpha in {0.2, 0.3, 0.4, 0.5}; wd in {0.0, 0.1, 1.0}	108	2,520	5.71 GB	f8dbac6a32658fb4

mechanism/ — mechanism

The boundary cells and the setup-D internals the mechanism analysis reads. Sept 2026 draft: Fig 7, Fig A1.

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
mechanism/checkpoints_boundary.tar	boundary	A	K in {20, 50, 97}; partition in {iid, operand}	20	1,200	4.12 GB	a3ee84ea4d64c4bd
mechanism/checkpoints_d_internals.tar	d_internals	D	alpha in {0.2, 0.3, 0.4, 0.5, 0.6, 0.25, 0.35, 0.45, 0.55, 0.225, 0.275, 0.325, 0.375, 0.425, 0.475, 0.525, 0.575}	85	11,200	4.15 GB	b5c76231ae52a849

legacy/ — legacy

checkpoints_ungrouped.tar holds 11 partial aggregation_alpha2 runs (old fede_addition_* ids, 2-15 checkpoints each) that are not in the run table. They were superseded by the runs in num_clients/checkpoints_aggregation_alpha2.tar and are kept only so the published record is unchanged. Do not use them for analysis.

archive	group	setups	what varies	runs	files	size	sha256 (first 16)
legacy/checkpoints_ungrouped.tar	ungrouped		-	11	128	0.07 GB	2715c0df2318d0d3

What is not here

• Strategy comparison (FedAvg, FedAdam, FedYogi, SCAFFOLD, FedAvgM, FedProx; Fig 5 and Fig A2 of the draft): those 90 runs were never checkpointed. Their histories are in the source repository.
• Setup A on the local-epochs axis (Fig 2, the anchor panel): the 12 runs at E in {1, 10, 25, 50} were launched with checkpointing off. Histories only.
• Centralised anchors outside the aggregation and d_internals groups, and every hyper-parameter calibration group: histories only.
• Per-round training histories and run specs for every run: committed in the source repository under results/runs/<run_id>/, not duplicated here.

What is in an archive

Paths inside each tar are results/runs/<run_id>/checkpoints/*.pt, holding:

prefix	what it is
ckpt_*.pt	global model state_dict at that round or epoch
client_*.pt	per-client weight signature — the channel the mechanism analysis reads
spectrum_*.pt	saved Fourier spectrum at that step

Verify and extract

sha256sum -c SHA256SUMS                          # from this directory
tar -xf heterogeneity/checkpoints_partitions.tar  # restores results/runs/<id>/checkpoints/

With huggingface_hub, one axis at a time:

from huggingface_hub import snapshot_download
snapshot_download("FedGrok/fedgrok-checkpoints", repo_type="dataset",
                  allow_patterns=["heterogeneity/*", "*.csv", "SHA256SUMS"])

Campaigns

campaign	archives	runs	size
v2 campaigns to 26 Aug 2026	12	520	34.2 GB
2-4 Sep 2026 sweep	4	144	5.7 GB