Skip to content

Releases

Spatial student checkpoints: LLS48 B224/B384 and CholecT45 L224

Choose a tag to compare

CurConMix spatial student checkpoints

These spatial encoder weights are separate from the MRTT temporal weights.

  • LLS48 SwinB(224): folds 0–4.
  • LLS48 SwinB(384): folds 0–4.
  • CholecT45 SwinL(224): folds 0–4.
  • Exactly 16 assets: 15 model-only files and spatial_sha256sums.txt.
  • Asset license: CC BY-NC 4.0; third-party rights remain applicable.

Download and loading instructions

Use a fresh clone of the current main branch for the activated download
manifests. GitHub's automatically generated source archive for this tag contains
the validated pre-publication source snapshot, whose download manifests are
still disabled. The checkpoint assets listed below are public.

All server SHA-256 digests and sizes match the verified exports. Each source
checkpoint's embedded validation predictions exactly match the corresponding
canonical stored predictions at original float32 precision. Exported model
tensors are unchanged. All 15 exports passed restricted and strict architecture
loading in their variant-specific runtime without forward passes. Optimizer
states, prediction arrays, private paths and training arguments are excluded.

These checks establish stored-checkpoint pairing and loadability, not new
inference or reproduction of published AP/F1. A source-derived keyed LLS48
B224/B384 feature exporter is provided, but image-based numerical qualification
and complete image-to-feature-to-MRTT reproduction remain outstanding.

LLS48 L224 spatial weights are unavailable. No new TMI-qualified CholecT45
B224/B384 set or feature arrays are included. Existing MICCAI resources are
preserved at MIDAS-SurgAI/CurConMix.

LLS48 consists of independent five-frame clips. Segmental evaluation remains
clip-bounded, resetting event extraction at every clip boundary; no continuity
across unsampled gaps or dense transition-localization claim is implied.

Research use only; these are not clinically validated medical devices.

LLS48 dataset and MRTT checkpoints v1.0.0

Choose a tag to compare

@yongyong98 yongyong98 released this 12 Oct 06:39

LLS48 dataset and MRTT checkpoints

This release provides 9,560 original annotated images in 1,912 independent
five-frame clips from 48 procedures, unchanged annotations and canonical
five-fold splits, and 15 LLS48 MRTT model-only checkpoints (B224, B384 and L224,
folds 0–4). The dataset uses 26 ZIP files; the release contains exactly 42 assets,
including the 15 weights and lls48_sha256sums.txt.

All asset sizes and server SHA-256 digests match the verified local exports.
Dataset ZIP CRC, exact membership and per-member hashes passed. MRTT exports
preserve model tensors and pass restricted and strict loading without forward
passes. Assets use CC BY-NC 4.0; source code uses MIT, with third-party terms
retained.

Use a fresh clone of the current main branch for the activated download
manifests. GitHub's automatically generated source archive for this tag contains
the validated pre-publication source snapshot, whose download manifests are
still disabled. The data and checkpoint assets listed below are public.

Spatial encoders are not included in this dataset/MRTT bundle. LLS48 B224/B384
spatial weights are available in the separate release; LLS48 L224 spatial weights
remain unavailable. Aligned features are required for MRTT. The source-derived
feature exporter has software checks but is not yet numerically qualified;
legacy feature arrays are not distributed. Complete image-to-feature-to-MRTT
paper-score reproduction has not been performed.

LLS48 segmental evaluation is clip-bounded: event extraction resets at each
independent five-frame clip. It measures local temporal consistency, not dense
transition localization or continuity across unsampled gaps. No new AP/F1,
training, model inference or forward pass was performed for this release.

Research use only; these are not clinically validated medical devices.