spacetime1008/diffcond-matched-control-audit

Matched-control audit of naïve difficulty conditioning: AdaPoinTr-family checkpoints

Checkpoints for Naïve Difficulty Conditioning Fails on ShapeNet-55: A Matched-Control Audit (Wenbo Gong, Transactions on Machine Learning Research, 2026).

• Paper and reviews: https://openreview.net/forum?id=cpe5PHAgoQ
• Code, configurations, logs and result artefacts: https://github.com/grenbel/diffcond-matched-control-audit

These are the 26 retained AdaPoinTr-family checkpoints listed with their SHA-256 in the supplementary material (Table 14). The files are byte-identical to that manifest; checkpoints.sha256 gives the digest of every file under its path here.

Files

File in the Hugging Face repository	Role in the paper	Seed	Epoch	Configuration (cfgs/ShapeNet55_models/)	Size (MB)
baseline_300ep/ckpt-best.pth	Baseline (300-epoch from-scratch AdaPoinTr; warm-start source)	–	best	AdaPoinTr.yaml	389.5
vanilla_ft/seed42/ckpt-epoch-030.pth	Vanilla-FT (matched control)	42	30	AdaPoinTr_baseline_smoke30_diffcrop.yaml	389.5
vanilla_ft/seed43/ckpt-epoch-030.pth	Vanilla-FT (matched control)	43	30	AdaPoinTr_baseline_smoke30_diffcrop.yaml	389.5
vanilla_ft/seed44/ckpt-epoch-030.pth	Vanilla-FT (matched control)	44	30	AdaPoinTr_baseline_smoke30_diffcrop.yaml	389.5
diffcond_full/seed42/ckpt-epoch-030.pth	DiffCond (full)	42	30	AdaPoinTr_DiffCond_smoke30.yaml	396.9
diffcond_full/seed43/ckpt-epoch-030.pth	DiffCond (full)	43	30	AdaPoinTr_DiffCond_smoke30.yaml	396.9
diffcond_full/seed44/ckpt-epoch-030.pth	DiffCond (full)	44	30	AdaPoinTr_DiffCond_smoke30.yaml	396.9
ablation_v1_lambda0/seed42/ckpt-epoch-030.pth	Ablation v1-λ0 (CE off, ranking on)	42	30	AdaPoinTr_DiffCond_v1_lambda0.yaml	395.9
ablation_v1_lambda0/seed43/ckpt-epoch-030.pth	Ablation v1-λ0 (CE off, ranking on)	43	30	AdaPoinTr_DiffCond_v1_lambda0.yaml	395.9
ablation_v1_lambda0/seed44/ckpt-epoch-030.pth	Ablation v1-λ0 (CE off, ranking on)	44	30	AdaPoinTr_DiffCond_v1_lambda0.yaml	395.9
ablation_v2b/seed42/ckpt-epoch-030.pth	Ablation v2b (CE off, ranking off)	42	30	AdaPoinTr_DiffCond_v2b.yaml	395.9
ablation_v2b/seed43/ckpt-epoch-030.pth	Ablation v2b (CE off, ranking off)	43	30	AdaPoinTr_DiffCond_v2b.yaml	395.9
ablation_v2b/seed44/ckpt-epoch-030.pth	Ablation v2b (CE off, ranking off)	44	30	AdaPoinTr_DiffCond_v2b.yaml	395.9
hard_specialist/seed42/ckpt-epoch-030.pth	Hard-Specialist	42	30	AdaPoinTr_hard_specialist.yaml	389.5
hard_specialist/seed43/ckpt-epoch-030.pth	Hard-Specialist	43	30	AdaPoinTr_hard_specialist.yaml	389.5
hard_specialist/seed44/ckpt-epoch-030.pth	Hard-Specialist	44	30	AdaPoinTr_hard_specialist.yaml	389.5
easywide_specialist/seed42/ckpt-epoch-030.pth	EasyWide-Specialist	42	30	AdaPoinTr_easywide_specialist.yaml	389.5
easywide_specialist/seed43/ckpt-epoch-030.pth	EasyWide-Specialist	43	30	AdaPoinTr_easywide_specialist.yaml	389.5
easywide_specialist/seed44/ckpt-epoch-030.pth	EasyWide-Specialist	44	30	AdaPoinTr_easywide_specialist.yaml	389.5
from_scratch_30ep_rerun/seed42/ckpt-best.pth	From-scratch 30-epoch baseline rerun (Pitfall 2 audit)	42	best	AdaPoinTr_baseline_smoke30.yaml	389.5
from_scratch_30ep_rerun/seed42/ckpt-epoch-029.pth	From-scratch 30-epoch baseline rerun (Pitfall 2 audit)	42	29	AdaPoinTr_baseline_smoke30.yaml	389.5
from_scratch_30ep_rerun/seed42/ckpt-epoch-030.pth	From-scratch 30-epoch baseline rerun (Pitfall 2 audit)	42	30	AdaPoinTr_baseline_smoke30.yaml	389.5
from_scratch_30ep_rerun/seed42/ckpt-last.pth	From-scratch 30-epoch baseline rerun (Pitfall 2 audit)	42	last	AdaPoinTr_baseline_smoke30.yaml	389.5
non_paper/vanilla_stock_crop/ckpt-epoch-030.pth	Not used in the paper: early vanilla fine-tune, stock crop path	–	30	AdaPoinTr_baseline_smoke30.yaml	389.5
non_paper/easy_specialist_pilot/ckpt-epoch-030.pth	Not used in the paper: Easy-Specialist pilot	–	30	AdaPoinTr_easy_specialist.yaml	389.5
non_paper/median_specialist_pilot/ckpt-epoch-030.pth	Not used in the paper: Median-Specialist pilot	–	30	AdaPoinTr_median_specialist.yaml	389.5

The three checkpoints in non_paper/ are not used for any number in the paper. The UpTrans and PCN checkpoints of the cross-backbone experiments were not retained and are not here.

Use

• Each file is a PyTorch checkpoint dictionary with the keys base_model (weights), optimizer, epoch, metrics and best_metrics; it loads with torch.load(path, map_location="cpu", weights_only=True).
• DiffCond and the two ablation conditions use models/AdaPoinTr_DiffCond.py from the GitHub repository; all other checkpoints use upstream PoinTr's models/AdaPoinTr.py. The configuration column names the file in the GitHub repository's cfgs/ShapeNet55_models/ that each checkpoint was trained with, and its REPRO.md explains how to evaluate or fine-tune with upstream PoinTr at commit 3f98676aeaabd9c1150dd04cd0a51f504949ab9c.
• Every fine-tuned checkpoint here was warm-started from baseline_300ep/ckpt-best.pth; baseline_300ep/ and from_scratch_30ep_rerun/ were trained from scratch.
• The metrics and best_metrics stored in the files are training-time validation values under PoinTr's random-crop validation proxy, not the paper's S/M/H results.

Verify a download

hf download spacetime1008/diffcond-matched-control-audit --local-dir diffcond-ckpts
cd diffcond-ckpts && sha256sum -c checkpoints.sha256

Licence

CC BY-NC 4.0. The models were trained on ShapeNet-55, which is derived from ShapeNet; using them must also comply with ShapeNet's terms of use (https://huggingface.co/spaces/ShapeNet/README/blob/main/TERMS.md), which allow non-commercial research and educational use only.

Citation

@article{gong2026naive,
  title   = {Na{\"i}ve Difficulty Conditioning Fails on {ShapeNet-55}: A Matched-Control Audit},
  author  = {Gong, Wenbo},
  journal = {Transactions on Machine Learning Research},
  year    = {2026},
  url     = {https://openreview.net/forum?id=cpe5PHAgoQ}
}