2,285 canonical result rows; 21 checkpoint archives; 57.6 GB of checkpoint payload. The archives contain 945 run snapshots: 931 canonical completed runs, 11 legacy partial runs, and three separately preserved CaMLSys reruns. Runs that did not save checkpoints are represented in the run table and source histories.

This update adds the completed cross-setup algorithm comparison, higher-local-epoch checkpoints, and the second-training-fraction aggregation supplements. All 17 previous archives and all previously published result-table values are preserved.

runs_v2.csv: configuration and outcome for each canonical run ID.
MANIFEST.csv: archive paths, groups, run/file counts, checkpoint payload sizes, SHA-256 checksums, and experiment axes.
SHA256SUMS: checksums of the complete tar files.
RELEASE_2026-09-21.json: exact run membership and extraction roots of the additions.
Source repository: code, specifications, per-round histories, and analysis.
Detailed release notes.

Access is gated with automatic approval. Request access, then use an authenticated download. Archives are uncompressed tar files. No previously released archive is renamed or replaced; a group can now have multiple supplement archives.

Archive inventory

Path	Group	Run snapshots	Checkpoint files	Payload GB
`num_clients/checkpoints_aggregation.tar`	aggregation	174	13,740	11.801
`num_clients/checkpoints_aggregation_alpha2.tar`	aggregation_alpha2	24	480	0.777
`num_clients/checkpoints_setup_k_ladder.tar`	setup_k_ladder	39	780	2.427
`num_clients/checkpoints_b_k20_wd_control.tar`	b_k20_wd_control	3	60	0.079
`num_clients/checkpoints_k_collapse_budget.tar`	k_collapse_budget	11	280	0.508
`local_epochs/checkpoints_local_epochs.tar`	local_epochs	36	720	1.666
`local_epochs/checkpoints_e50_long.tar`	e50_long	6	120	0.194
`participation/checkpoints_participation.tar`	participation	9	540	0.626
`participation/checkpoints_participation_setups.tar`	participation_setups	30	780	1.095
`heterogeneity/checkpoints_dirichlet_band.tar`	dirichlet_band	24	1,440	2.523
`heterogeneity/checkpoints_dirichlet_setups.tar`	dirichlet_setups	72	1,590	2.770
`heterogeneity/checkpoints_size_control.tar`	size_control	12	720	1.424
`heterogeneity/checkpoints_partitions.tar`	partitions	108	2,520	5.709
`mechanism/checkpoints_boundary.tar`	boundary	20	1,200	4.124
`mechanism/checkpoints_d_internals.tar`	d_internals	85	11,200	4.151
`legacy/checkpoints_ungrouped.tar`	ungrouped	11	128	0.068
`algorithms/checkpoints_h2_mechanism.tar`	h2_mechanism	38	2,280	1.532
`algorithms/checkpoints_algo_comparison.tar`	algo_comparison	189	5,580	13.220
`local_epochs/checkpoints_local_epochs_supplement_20260921.tar`	local_epochs	18	840	1.797
`num_clients/checkpoints_aggregation_alpha2_supplement_20260921.tar`	aggregation_alpha2	33	960	1.013
`num_clients/checkpoints_aggregation_alpha2_reruns_20260921.tar`	aggregation_alpha2	3	60	0.079

Extraction and provenance

Canonical checkpoints extract to results/runs/<run_id>/checkpoints/ and join to runs_v2.csv by run ID. Histories and specifications live in the source repository. The source distinguishes first threshold crossing, sustained crossing, memorization, and final accuracy; these are not interchangeable measures.

The three reruns extract to results/reruns/camlsys-2026-09-21/<run_id>/checkpoints/. Their matching result.json, specification, and history are in that source directory. Do not pair these weights with the canonical CSV outcome or count them as new seeds. The original weights for fede_addition_3faf5207eea4, fede_addition_4dc6abeee354, and fede_addition_e9de3f5a209a are unavailable in this release; their published result rows remain unchanged. The separate rerun archive is not a replacement for them.

The 11 snapshots in legacy/checkpoints_ungrouped.tar are historical partial runs, not part of the canonical run table. Exclude them from analyses.

Five longer weight-decay runs were still in progress at collection and are excluded from this snapshot. The S5-transformer algorithm calibration remains outstanding.

sha256sum -c SHA256SUMS
tar -xf algorithms/checkpoints_algo_comparison.tar

With huggingface_hub, download one axis at a time:

from huggingface_hub import snapshot_download
snapshot_download(
    "FedGrok/fedgrok-checkpoints",
    repo_type="dataset",
    allow_patterns=["algorithms/*", "*.csv", "SHA256SUMS", "RELEASE_2026-09-21.json"],
)

Downloads last month: 69

Total file size:

57.6 GB