Document training data on the model card?
The model card currently says training datasets for the released v1 weights are "not publicly documented," while the AutowareFoundation/bevfusion card explicitly states its models were trained on TIER IV's internal database.
Could the training data for the camera_streampetr v1 artifacts be confirmed and
documented on the card as well? Since pretrained weights generally inherit the
license terms of their training data, this matters for anyone doing license-compliance review before adopting the model — e.g., whether any public dataset with non-commercial terms (nuScenes etc.) was involved. The 5-camera layout and class set suggest an internal dataset, but having it stated on the card would remove the ambiguity.
Thanks for publishing these — the bevfusion card's provenance section is a great example of what makes adoption decisions easy.
The model card currently says training datasets for the released v1 weights are "not publicly documented," while the AutowareFoundation/bevfusion card explicitly states its models were trained on TIER IV's internal database.
Hi @haroldd , thanks for raising this issue. I don't see a contradiction here, they were trained on TIER IV's internal database and not publicly documented.
But we can ask the TIER IV team if any non-commercial datasets like nuScenes were involved and if model weights are compliant with the stated license.
@ossyaritoori could you check this with the internal TIER IV team and provide help with this issue?
Thanks! any update here?