BiRefNet Portrait: exact checkpoint lineage and validation splits

[BoazTelem]

We are evaluating BiRefNet Portrait for commercial portrait cutouts and want to
record reproducible checkpoint provenance alongside its MIT license.

We saw your code/weight licensing clarification
and dataset licensing discussion.
The remaining questions concern the exact Portrait training history.

The artifact is ZhengPeng7/BiRefNet-portrait
at revision b6561965a70070d9143fd9e558f6ca3c481510db, file
model.safetensors. Hugging Face's LFS metadata reports SHA-256
4a4eb3a5469b75f0cccaec6772c22fc30e6c12ed429c2c0ba71b43e3d8d97182.

Could you point us to records covering these points for that exact checkpoint?

1. Initialization and resume history: the training code revision, run config,
   and each initialization/resume checkpoint's source and digest. Did it start
   from a pretrained Swin backbone only, or inherit another BiRefNet checkpoint?
   The exported bb_pretrained: false setting describes loading, so we could
   not infer the original training history from it.
2. Dataset versions and splits: the card lists P3M-10k excluding P3M-500-P,
   plus TR-humans. Does that include P3M-500-NP in training? Was P3M-500-P used
   to select the epoch, loss settings or other hyperparameters, in addition to
   reporting validation scores? Were any other data used by this checkpoint or
   its parent checkpoints?
3. Synthetic-data identity: which TR-humans snapshot/archive was used,
   ideally identified by a digest or file list? Is there a record of its
   generator checkpoints and background-image sources?
4. Applicable notices: are there checkpoint-specific notices or permissions
   covering the initialization weights, training sources and copied model
   components that should accompany the MIT code/weight license? Links to
   existing rights/provenance documentation would be helpful.

Thank you. A training config, log or model-card update addressing these points
would let us distinguish documented provenance from assumptions.

[ZhengPeng7]

Of course the license covers all the latest weights files and source codes.
Check the model cards on HF or the model zoo section in the GitHub repo's README for training sets.
The author of the humans dataset set it as invisible. I'm not sure what the problem. You can send him some emails to check it. I surely hope it can be resumed publicly.

[BoazTelem]

Thank you for confirming that the license covers the current code and weights.
We checked the Portrait model card and model-zoo row, and will direct the humans
archive/generation questions to its author.

For the exact Portrait checkpoint linked in the issue, could you clarify two
remaining points that we could not find there?

1. Was it initialized from a pretrained Swin backbone only, or resumed/fine-tuned
   from another BiRefNet checkpoint? The initialization checkpoint name or an
   existing training config/log would help us record the inherited data exposure.
2. Does its P3M-10k training set include P3M-500-NP? Was P3M-500-P used to select
   the epoch or hyperparameters, or only to report results?

If the humans snapshot used for this run has a date, file count or archive hash,
that would also help us match the dataset author's records. We are not requesting
private images or a new training run. Thank you.

[ZhengPeng7]

For your questions here:

1. Only official bb weights provided by the Swin Transformer authors.
2. The datasets used for validation were never included in the training. Regarding hyperparameter selection, theoretically, they can be used for that, but I didn't adapt them for the portrait task.