LaUER checkpoints

LaUER learns a fused ECG signal–image representation for disease detection and interpretation using clinical language at three granularities. The stages are cumulative.

Checkpoint suffix	Stage	Components added
SGL	Semantic Granularity Learning	Signal, image, and clinical text alignment
SGL_CGL	Concept Granularity Learning	Signal–image fusion, matching, and query-guided tracing
SGL_CGL_TGL	Token Granularity Learning	Autoregressive report and explanation generation

See checkpoints.json for the checkpoint and matching YAML paths, training seed, and epoch recorded in each artifact. Seed folder names group the source runs. Use the index's training_seed to identify the seed stored in a specific checkpoint. The seed20260915 CGL checkpoint is the _o variant and has a matching YAML configuration.

Loading

Use the anonymous LaUER implementation with the YAML paired to each checkpoint. Run commands from the code repository root. Follow its setup instructions to obtain the frozen MedCPT Article Encoder and MedCPT Query Encoder in the weights/ directory.

Example using the seed-42 TGL artifact after downloading this repository to release/:

python inference.py   --config release/configs/seed42/LaUER_SGL_CGL_TGL.yaml   --checkpoint release/checkpoints/seed42/LaUER_SGL_CGL_TGL.pt   --signal /path/to/record.mat   --image /path/to/record.png   --device cuda

For zero-shot evaluation, pass the paired checkpoint and YAML to python -m zeroshot.run_experiments after preparing the benchmark datasets.

Integrity

SHA256SUMS lists the weights and configurations. Every checkpoint was reloaded against its paired architecture and each parameter tensor was verified against the source tensor after renaming. These are research artifacts. The original training logs and metrics are outside the release.