Commit History
-
Rename the model to Lumen: registry key and checkpoint path in provenance
Verified signatureThe framework and the model were renamed from PathoVLM/odyssey to Lumen. adapter_provenance.json recorded the old registry key and the old training output path; modeling_lumen.py named the old package in its docstring. Weights are untouched. model.safetensors is byte-identical, sha256 3559c771a769286939a0916040e62a48cc25afe7c8bcd887ab3bebd5914c3c89, and the rebuilt model still reproduces the evaluated checkpoint exactly.b89e805
-
Refuse unexpected forward kwargs, and document the loading warning
Verified signatureforward() took **kwargs and dropped anything it did not recognise, so the conventional pixel_values=/input_ids= returned a dict of Nones with no error. It now raises naming the arguments it does take. The card gained a runnable inference example and an explanation of the 'newly initialized' warning, which is expected: the backbones are withheld by design and all 455 vision and 199 text tensors were verified identical to the public releases. Weights unchanged, sha256 3559c771...12ec91e -
Lumen alignment: rank-4 LoRA adapters, projection heads, logit scale
Verified signature2,984,961 parameters exported from the checkpoint the manuscript reports (registry key odyssey_retrain_quilt). The 740,721,664 frozen Virchow2 and BioMedBERT parameters are withheld and come from their own repositories. model.safetensors sha256 3559c771a769286939a0916040e62a48cc25afe7c8bcd887ab3bebd5914c3c892076468