Fine-tuned transformer checkpoints for a cross-architecture audit of attention-head-masking OOD detection (NeurIPS 2026)
Year: 2026
Type: dataset
Abstract: Fine-tuned transformer checkpoints accompanying the NeurIPS 2026 Evaluations and Datasets Track paper "Auditing Attention Head Masking for Out-of-Distribution Detection: Cross-Architecture Failures and Polarity Inversions" by Vishnu Bindu Balachandran: https://openreview.net/forum?id=bqTe5C6VVH Contents: 90 fine-tuned checkpoints (44.26 GB of weights) in 19 uncompressed ZIP archives, each holding complete runs. Seeds 42, 123, 456, 789, 1024. promised40_layoutlmv3_tobacco3482_intra.zip / _cross.zip — LayoutLMv3-base, Tobacco3482 (CC BY-NC-SA 4.0) promised40_layoutlmv3_financedocs.zip — LayoutLMv3-base, FinanceDocs intra / cross (CC BY-NC-SA 4.0) promised40_vit_b16_cifar10_intra.zip / _cifar10_svhn.zip / _cifar100.zip — ViT-B/16 (CC BY 4.0) promised40_roberta_base_sst2.zip — RoBERTa-base, SST-2 (CC BY 4.0) rebuttal_vit_b16_cifar100_c10.zip / _c20.zip / _c50.zip — ViT-B/16, CIFAR-100 class-count ablation (CC BY 4.0) rebuttal_roberta_large_sst2_part1of3.zip … part3of3.zip — RoBERTa-large, SST-2 (CC BY 4.0) rebuttal_nearood_vit_smallmammals.zip / rebuttal_nearood_roberta_20news_comp.zip — near-OOD fine-tunes (CC BY 4.0) camera_ready_layoutlmv3_tobacco3482_leaveout_Memo.zip / _Form / _Resume / _Scientific — LayoutLMv3-base, one Tobacco3482 class held out (CC BY-NC-SA 4.0) The 40 checkpoints in the promised40_* archives are listed in SHA256SUMS_promised40.txt, which is identical to manifests/SHA256SUMS.txt in the code repository. README.md explains how to verify and load the checkpoints; checkpoint_index.json lists every weight file with its archive, licence, size and SHA-256. LayoutLMv3 derivatives: CC BY-NC-SA 4.0 (inherited from microsoft/layoutlmv3-base). ViT-B/16, RoBERTa-base and RoBERTa-large derivatives: CC BY 4.0. Training data remain under their original licences and are not included. Code, per-seed results, statistics, attention-head masks, split configurations and Croissant metadata: https://github.com/VISHNUBINDUBALACHANDRAN/ood-evaluation-neurips2026
Author: Balachandran, Vishnu Bindu
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225328142