aidm-dogs-vs-cats-models

The final checkpoints of a dogs-vs-cats image-classification study. Each checkpoint is named by a record of results/runs_final.jsonl; the full experiment record, including every ablation run, lives in the companion dataset repo.

Generated by scripts/90_publish_hf.py on 2026-09-23 07:35 UTC. Every count, fingerprint and metric below was read from the run registry during that run, so this card describes the runs that exist, not the runs that once existed.

Checkpoints

file	used by run(s)	arch	seed
t_convnext_full-s0-7b21d385/best.pt	final_t_convnext_full, final_ensemble_top3	timm:convnext_tiny.fb_in1k	0
t_vit_b16_full-s0-35a4c1c8/best.pt	final_t_vit_b16_full, final_ensemble_top3	timm:vit_base_patch16_224.augreg_in21k_ft_in1k	0
t_vit_b16_llrd-s0-ddae88d4/best.pt	final_t_vit_b16_llrd, final_ensemble_top3	timm:vit_base_patch16_224.augreg_in21k_ft_in1k	0

Why best.pt and not last.pt

Each run directory holds two checkpoints. best.pt is the weights of the epoch with the best validation score, and it is the file every metric in the registry was computed from. last.pt is the final-epoch state plus the optimiser and scheduler state; it exists only to resume an interrupted run, no recorded number comes from it, and it would double the size of this repo. Only best.pt is published.

Runs in this snapshot

The registry holds 4027 distinct run(s) across 134 registry file(s) (9941 record(s) before de-duplication on (config_hash, seed, splits_fingerprint)).

By task

task	runs
cifar10	43
cifar10lt	195
dogcat	3789

By tier

tier	runs
ablation	3613
cifar	43
cifar_lt	195
foundation	30
retune	50
stability	24
transfer	50
transformer	22

Splits fingerprint

Every run record carries the fingerprint of the splits.json that produced its train / val / holdout indices. Results with different fingerprints are not comparable.

task	splits_fingerprint
cifar10	3d6f0150310f
cifar10lt	3d6f0150310f
dogcat	0c31c0f203df, 3138e184ef5f

Held-out results (results/runs_final.jsonl, split val_holdout)

run	acc	balanced_acc	auc	log_loss	ece	n
final_t_convnext_full	0.9975	0.9975	0.999969	0.057143	0.0502	2000
final_t_vit_b16_full	0.9970	0.9970	0.999882	0.060195	0.0511	2000
final_t_vit_b16_llrd	0.9965	0.9965	0.999529	0.060440	0.0497	2000
final_ensemble_top3	0.9975	0.9975	0.999957	0.058023	0.0500	2000

Package versions recorded in the runs

package	version(s)
numpy	1.26.4
timm	1.0.29
torch	2.8.0+cu128
torchvision	0.23.0+cu128

Files

• runs/<run_id>/best.pt -- the best-epoch checkpoint of every run in the registry, one file per run, in a directory named by the run_id that indexes that run's record. (2243 file(s), 68.4 GiB)
• README.md -- this card, generated from the registry at publish time.

Loading

Each file is a torch.save checkpoint. Rebuild the architecture named in the arch column with timm, then model.load_state_dict(torch.load(path)['model']). The exact training config of each run is in the matching record of results/runs_final.jsonl in the dataset repo.