Clarify model licensing scope and dataset provenance

Separate implementation licenses from pretrained-weight and training-data terms. Preserve existing valid grants and document unresolved distribution permissions. No weights or access settings changed.

LICENSE

@@ -0,0 +1,49 @@
+# face_multitask_v2 — licensing scope notice
+
+The weights have been designated for **noncommercial research**. Py-Feat's
+original implementation is MIT, but the initialized
+[`convnextv2_tiny.fcmae_ft_in22k_in1k` weights](https://huggingface.co/timm/convnextv2_tiny.fcmae_ft_in22k_in1k)
+are **CC BY-NC 4.0**, distinct from the backbone's MIT source code.
+
+For `face_multitask_v28.safetensors`, DISFA+ and EYEDIAP are excluded from
+gradient training; their benchmark scores were used in checkpoint selection.
+DISFA (without +) is training data. Training also includes BP4D, BP4D+, CK+,
+UNBC-McMaster PAIN, EmotioNet, AM-FED, Aff-Wild2, AffectNet, RAF-DB, FER+,
+ExpW, MELD, AFEW-VA, ETH-XGaze, Gaze360, MPIIGaze, Columbia Gaze, and
+CelebV-HQ across the training stages. Older checkpoints have different
+provenance and must be assessed separately.
+
+For the current v2.8 checkpoint, [Gaze360's research license, section 2(3)](https://github.com/erkil1452/gaze360/blob/master/LICENSE.md)
+expressly excludes commercial applications of models trained on its dataset.
+[ETH-XGaze](https://ait.ethz.ch/xgaze) access conditions expressly prohibit
+training models for commercial products. These are express restrictions,
+additional to the pretrained backbone's terms; they are not assumptions that
+all dataset licenses automatically transfer to weights. Gaze360's use and
+redistribution limits and ETH-XGaze's additional dataset/software terms also
+require separate review of public weight-sharing permission. The applicable
+acquisition terms and checkpoint provenance must be established separately
+for older files in this repository.
+
+See the [v2 licensing notice](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V2.md)
+and [dataset register](https://github.com/cosanlab/py-feat/blob/main/LICENSES/DATASETS.md).
+The research designation does not establish unrestricted public redistribution:
+the relevant dataset agreements, chronology, and any required permissions
+must be resolved for each checkpoint. This notice does not grant third-party
+rights or assume that all dataset or teacher terms automatically attach to
+every trained artifact. The optional ArcFace identity branch has its own
+terms and is not the basis for this network's licensing restrictions.
+
+This notice clarifies scope; it does not revoke existing valid grants or create a new license for third-party material. Dataset and teacher terms do not automatically relicense every trained artifact or inference output. A software license or model-card badge alone does not establish all checkpoint redistribution or commercial-use permissions.
+
+## Commercial-use inquiries
+
+These multitask weights are not offered as commercially cleared weights.
+Py-Feat does not broker third-party permissions or offer a commercial
+sublicense overriding upstream terms. Users can use independently cleared
+weights or contact the applicable rights holders directly for any necessary
+permissions covering the exact assets and commercial use. Dataset download
+approval alone is not necessarily such permission. See the
+[v2 commercial-use notice](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V2.md#commercial-use-inquiries)
+and [commercial retraining table](https://github.com/cosanlab/py-feat/blob/main/LICENSES/COMMERCIAL-RETRAINING.md).
+Existing valid grants remain unchanged; downstream permissions do not
+automatically resolve the distributor's separate obligations.

README.md

@@ -17,6 +17,8 @@ pipeline_tag: image-classification
 
 # face_multitask_v2
 
+> **Licensing scope:** See the [license and provenance notice](LICENSE) before relying on this card's license metadata for pretrained-weight redistribution or commercial use. Existing valid grants are preserved.
+
 A single multi-task convolutional model for facial behavior analysis, used by
 [py-feat](https://github.com/cosanlab/py-feat)'s `Detectorv2`. From one face crop
 it jointly predicts **action units, categorical emotion, valence/arousal,
@@ -146,9 +148,53 @@ The model expects a face crop produced by RetinaFace + py-feat's
 `extract_face_from_bbox_torch(frame, bbox, face_size=256, expand_bbox=1.2)`,
 then center-cropped to 224 and ImageNet-normalized. `Detectorv2` handles this.
 
-## License
-
-**Research / non-commercial use only.** Trained on datasets (AffectNet, DISFA+,
-RAF-DB, Aff-Wild2, AFEW-VA, BP4D, ETH-XGaze, EYEDIAP, etc.) whose licenses
-restrict use to academic research. The ConvNeXt-V2 backbone is MIT-licensed.
-Confirm each constituent dataset's terms before any non-research use.
+## License and provenance
+
+The weights have been designated for **noncommercial research**. Py-Feat's
+original implementation is MIT, but the initialized
+[`convnextv2_tiny.fcmae_ft_in22k_in1k` weights](https://huggingface.co/timm/convnextv2_tiny.fcmae_ft_in22k_in1k)
+are **CC BY-NC 4.0**, distinct from the backbone's MIT source code.
+
+For `face_multitask_v28.safetensors`, DISFA+ and EYEDIAP are excluded from
+gradient training; their benchmark scores were used in checkpoint selection.
+DISFA (without +) is training data. Training also includes BP4D, BP4D+, CK+,
+UNBC-McMaster PAIN, EmotioNet, AM-FED, Aff-Wild2, AffectNet, RAF-DB, FER+,
+ExpW, MELD, AFEW-VA, ETH-XGaze, Gaze360, MPIIGaze, Columbia Gaze, and
+CelebV-HQ across the training stages. Older checkpoints have different
+provenance and must be assessed separately.
+
+For the current v2.8 checkpoint, [Gaze360's research license, section 2(3)](https://github.com/erkil1452/gaze360/blob/master/LICENSE.md)
+expressly excludes commercial applications of models trained on its dataset.
+[ETH-XGaze](https://ait.ethz.ch/xgaze) access conditions expressly prohibit
+training models for commercial products. These are express restrictions,
+additional to the pretrained backbone's terms; they are not assumptions that
+all dataset licenses automatically transfer to weights. Gaze360's use and
+redistribution limits and ETH-XGaze's additional dataset/software terms also
+require separate review of public weight-sharing permission. The applicable
+acquisition terms and checkpoint provenance must be established separately
+for older files in this repository.
+
+See the [v2 licensing notice](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V2.md)
+and [dataset register](https://github.com/cosanlab/py-feat/blob/main/LICENSES/DATASETS.md).
+The research designation does not establish unrestricted public redistribution:
+the relevant dataset agreements, chronology, and any required permissions
+must be resolved for each checkpoint. This notice does not grant third-party
+rights or assume that all dataset or teacher terms automatically attach to
+every trained artifact. The optional ArcFace identity branch has its own
+terms and is not the basis for this network's licensing restrictions.
+
+This notice clarifies scope; it does not revoke existing valid grants or create a new license for third-party material. Dataset and teacher terms do not automatically relicense every trained artifact or inference output. A software license or model-card badge alone does not establish all checkpoint redistribution or commercial-use permissions.
+
+
+## Commercial-use inquiries
+
+These multitask weights are not offered as commercially cleared weights.
+Py-Feat does not broker third-party permissions or offer a commercial
+sublicense overriding upstream terms. Users can use independently cleared
+weights or contact the applicable rights holders directly for any necessary
+permissions covering the exact assets and commercial use. Dataset download
+approval alone is not necessarily such permission. See the
+[v2 commercial-use notice](https://github.com/cosanlab/py-feat/blob/main/LICENSES/PYFEAT-V2.md#commercial-use-inquiries)
+and [commercial retraining table](https://github.com/cosanlab/py-feat/blob/main/LICENSES/COMMERCIAL-RETRAINING.md).
+Existing valid grants remain unchanged; downstream permissions do not
+automatically resolve the distributor's separate obligations.