NOTICE: WEIGHTS LICENSING

No pretrained weight files are distributed in this directory or anywhere
else in the source tree. The conversion scripts in this directory are
part of the LibreYOLO source repository and are covered by the root
MIT License.

Pretrained LibreYOLO weights are published separately on Hugging Face
under the LibreYOLO organization:

    https://huggingface.co/LibreYOLO

Each Hugging Face model repository ships its own LICENSE, NOTICE (where
applicable), and attribution reflecting the license of the upstream
project the weights were derived from. Per-family summary:

    YOLOX    (LibreYOLOX{n,t,s,m,l,x})                 Apache-2.0
             upstream: Megvii-BaseDetection/YOLOX

    YOLO9    (LibreYOLO9{t,s,m,c})                     MIT
             upstream: MultimediaTechLab/YOLO

    YOLOv1   (LibreYOLO1{t,b})                         public domain
    YOLOv2   (LibreYOLO2{t,b})                         public domain
    YOLOv3   (LibreYOLO3{t,b,spp})                     public domain
    YOLOv4   (LibreYOLO4{t,b})                         public domain
             upstream: pjreddie/darknet (v1/v2/v3), AlexeyAB/darknet (v4)
             Darknet is public domain ("YOLO LICENSE"); both the .cfg
             architectures and the pretrained .weights carry no license
             obligations. Convert with weights/convert_darknet_weights.py.
             YOLOv1 is Pascal VOC (20 classes), fixed 448. Its pretrained
             yolov1.weights was pulled from pjreddie.com/media/files/ (now
             archived only on the Wayback Machine); the tiny-yolov1 .weights
             was overwritten upstream by the YOLOv2 release and is not
             recoverable, so LibreYOLO1t ships as code only (bring your own
             tiny-yolov1.weights to convert).

    Dome-DETR (LibreDOMEDETR{s,m,l}-{aitod,visdrone})  NOT REDISTRIBUTED
             upstream: https://huggingface.co/RicePasteM/Dome-DETR
             Not mirrored to the LibreYOLO organization. The upstream model
             card carries no license field in its metadata, and its prose
             simultaneously states the project is Apache-2.0 and that the
             material is "for academic research purposes only". Those do not
             agree, and the stricter reading is not a redistribution grant, so
             the weight license is implied from the source repository rather
             than stated. Until upstream sets an explicit license, download the
             checkpoints from the link above and convert locally with
             weights/convert_domedetr_weights.py. This follows the YOLO-NAS
             precedent of linking rather than rehosting.
             There is no COCO checkpoint for this family: only AI-TOD-V2
             (9 classes) and VisDrone (12), hence the dataset suffix on every
             canonical filename and the absence of a bare LibreDOMEDETRs.pt.

    YOLOv7   (LibreYOLO7b)                             MIT
             upstream: MultimediaTechLab/YOLO (Kin-Yiu Wong & Hao-Tang Tsui)
             The authors' MIT re-release, not WongKinYiu/yolov7 (GPL-3.0).
             Convert v7.pt with weights/convert_yolo7_weights.py.

    RF-DETR  (LibreRFDETR{n,s,m,l})                    Apache-2.0
             upstream: roboflow/rf-detr
             (backbone: facebookresearch/dinov2, Apache-2.0)

    Faster R-CNN (LibreFasterRCNN{n,s,m,l})            BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreFasterRCNN{n,s,m,l}

    RetinaNet (LibreRetinaNet{r50,r50v2})               BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreRetinaNet{r50,r50v2}
    SSD300   (LibreSSD300)                            BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoint carries no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project and
             repeats torchvision's pretrained-model/data warning. The VGG-16
             backbone initialization traces to Oxford Visual Geometry Group
             feature weights by Karen Simonyan and Andrew Zisserman under
             CC BY 4.0; the model NOTICE preserves attribution and changes.
             Public mirror: https://huggingface.co/LibreYOLO/LibreSSD300
    Mask R-CNN (LibreMaskRCNNr50)                      BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             official checkpoint:
               maskrcnn_resnet50_fpn_v2_coco-73cbd019.pth
             official SHA-256:
               73cbd0190fcbe3ba339921fbce2c3a0b6bb9126c9a133c85e43a2a8e060a109e
             The checkpoint carries no per-object license file. Rehosting
             uses BSD-3-Clause implied by the releasing project, not a
             publisher-confirmed checkpoint grant. The weight repository
             ships the verbatim BSD text and torchvision pretrained-model
             caveat. Conversion adds metadata without changing learned tensors.
             Public mirror: https://huggingface.co/LibreYOLO/LibreMaskRCNNr50
    FCN      (LibreFCN{r50,r101})                      BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreFCN{r50,r101}
    CenterNet (LibreCenterNet{resdcn18,dla34})          MIT (implied)
             upstream code and official COCO checkpoints:
               xingyizhou/CenterNet at commit
               4c50fd3a46bdf63dbf2082c5cbb3458d39579e6c (MIT)
             Official checkpoint objects carry no standalone license file.
             Rehosting uses MIT implied by the releasing project and the
             maintainer-approved port brief, not a publisher-confirmed
             checkpoint-specific grant. Every mirror ships the verbatim
             CenterNet MIT license and repeats this caveat.
             resdcn18 source SHA-256:
               f9e413f91cdb235adbcb41c5c4052b8f7ff53999374048949789c29d6df18eaa
             dla34 source SHA-256:
               43bf4cc2efe00e02c1ae8484035b062a35543872d276c7dcfeb4db3e64203e4f
             Public mirrors: https://huggingface.co/LibreYOLO/LibreCenterNet{resdcn18,dla34}
    FCOS       (LibreFCOSr50)                           BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoint carries no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. The weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case.
             Public mirror: https://huggingface.co/LibreYOLO/LibreFCOSr50
    AlexNet  (LibreAlexNetb-cls)                       BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official ImageNet-1K checkpoint has no per-object license
             file. Rehosting uses BSD-3-Clause implied by the releasing
             project, not a publisher-confirmed checkpoint grant. The weight
             repo ships the verbatim BSD text and states torchvision's warning
             that pretrained-model terms may derive from training data.
             Public mirror: https://huggingface.co/LibreYOLO/LibreAlexNetb-cls
    VGG      (LibreVGG{16,19,16bn,19bn}-cls)           BSD-3-Clause (implied)
             upstream code: pytorch/vision at commit
               10f68dbd78b9aa5cab9328f3b2e99cfb0b608122 (BSD-3-Clause)
             The official ImageNet-1k V1 checkpoints carry no per-object
             license file. Rehosting uses BSD-3-Clause implied by the
             releasing project, not a publisher-confirmed checkpoint grant.
             Every weight repo ships the verbatim BSD text and states
             torchvision's warning that pretrained-model terms can depend on
             training data and users must determine permission for their use
             case. Conversion leaves every learned tensor unchanged and only
             adds LibreYOLO checkpoint metadata.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreVGG{16,19,16bn,19bn}-cls
    DeepLabv3 (LibreDeepLabv3{r50,r101,mv3}-sem)        BSD-3-Clause (implied)
             upstream code: pytorch/vision v0.26.0 at commit
               336d36e8db990a905498c73933e35231876e28bc (BSD-3-Clause)
             The official COCO checkpoints carry no per-object license file.
             Rehosting uses BSD-3-Clause implied by the releasing project,
             not a publisher-confirmed checkpoint grant. Every weight repo
             ships the verbatim BSD text and states torchvision's warning that
             pretrained-model terms may derive from training data and users
             must determine permission for their use case. COCO annotations
             are CC BY 4.0; source images retain individual Flickr terms.
             Public mirrors: https://huggingface.co/LibreYOLO/LibreDeepLabv3{r50,r101,mv3}-sem
    EfficientDet (LibreEfficientDet{d0,d1,d2,d3,d4})  Apache-2.0 (implied)
             upstream: rwightman/efficientdet-pytorch 0.4.1 at commit
               c6dff775a36cea0bf9b76c58e59f936411c5ce01
             Official D0-D4 COCO checkpoints come from the upstream v0.1
             GitHub release and retain every learned tensor unchanged. The
             release assets carry no separate weight-license object, so
             rehosting uses Apache-2.0 implied by the releasing project rather
             than a publisher-confirmed checkpoint-specific grant. Every
             weight repository ships the verbatim Apache-2.0 text and a NOTICE
             stating this caveat.

    Deformable DETR                                     Apache-2.0
             (LibreDeformableDETR{r50ss,r50ssdc5,r50,
              r50refine,r50twostage})
             upstream code: fundamentalvision/Deformable-DETR
             upstream weights: SenseTime/deformable-detr* on Hugging Face
             Converted without changing learned tensors; the 91-column COCO
             head is retained and mapped to contiguous COCO-80 at runtime.

    DINO-DETR (LibreDINODETR{r50,r50s5,swinl})          Apache-2.0 (implied)
             upstream code and weight release: IDEA-Research/DINO at commit
               d84a491d41898b3befd8294d1cf2614661fc0953
             Official files: checkpoint0011_4scale.pth,
               checkpoint0011_5scale.pth, and
               checkpoint0027_5scale_swin.pth from the authors' Google Drive.
             The releasing repository declares Apache-2.0, but its checkpoint
             files have no separate license file or metadata. Rehosting uses
             Apache-2.0 implied by that repository declaration, not a
             publisher-confirmed checkpoint grant. Every mirror ships the
             verbatim repository license, an explanatory NOTICE, source and
             converted SHA-256 hashes, and an exact provenance record. Learned
             tensors are unchanged; conversion only adds checkpoint metadata.
             See docs/provenance/dinodetr.md and convert with
             weights/convert_dinodetr_weights.py.

    RT-DETR  (LibreRTDETR{r18,r34,r50,r50m,r101})      Apache-2.0
             upstream: lyuwenyu/RT-DETR
             (backbone: PaddlePaddle/PaddleClas ResNet_vd, Apache-2.0)

    RT-DETRv2 OBB                                      Apache-2.0
             (LibreRTDETRv2{n,s,m,l,x}-obb)
             upstream code: RicePasteM/RiO-DETR at commit
               22d5232a4e0df6ac4bc26ed1c8aac8b4060449c7
             official DOTA 1.0 checkpoints: RicePasteM/RT-DETR-OBB at
               revision f376e9dcedfb9a47a21ac71ef61ad99f8b545698
             converted mirrors: LibreYOLO/LibreRTDETRv2{n,s,m,l,x}-obb
             Each mirror carries the upstream Apache License 2.0 and NOTICE.
             Convert a locally supplied official checkpoint with
             weights/convert_rtdetrv2_weights.py; conversion preserves learned
             tensors and adds only LibreYOLO checkpoint metadata.

    DETR     (LibreDETR{r50,r50dc5,r101,r101dc5})      Apache-2.0
             upstream: facebookresearch/detr, commit
             29901c51d7fe8712168b8d0d64351170bc0f83e0.
             Official COCO model-zoo tensors are unchanged; conversion only
             adds LibreYOLO checkpoint metadata. Every Hugging Face weight
             repository carries the upstream Apache-2.0 LICENSE and NOTICE.

    EoMT     (LibreEoMTl-sem, LibreEoMTl-seg,          MIT
              LibreEoMTl-seg-1280, LibreEoMT{s,b,l}-panoptic)
             upstream: tue-mps/eomt
             (backbone: facebookresearch/dinov2, Apache-2.0)
             Panoptic checkpoints load as first-class task="panoptic"
             (133 COCO classes: 80 things + 53 stuff).

    MobileNetV4 (LibreMobileNetV4{s,m,l}-cls)          Apache-2.0
             upstream: huggingface/pytorch-image-models (timm), ImageNet-1k

    ConvNeXt (LibreConvNeXt{t,s,b}-cls)                Apache-2.0 (weights)
             upstream: timm fb_in1k; arch facebookresearch/ConvNeXt (MIT)
             (ConvNeXt-V2 fcmae weights are CC-BY-NC and are NOT used)

    DeiT     (LibreDeiT{t,s,b}-cls)                    Apache-2.0
             upstream weights: timm *.fb_in1k at the immutable revisions and
             SHA-256 digests recorded in docs/provenance/deit.md
             architecture: facebookresearch/deit and timm (Apache-2.0)
             Only plain patch-16 224px checkpoints are mirrored; distilled,
             DeiT III, and 384px variants are NOT used.

    EfficientNetV2 (LibreEfficientNetV2{b0,b1,b2,b3}-cls)  Apache-2.0
             upstream: timm (.in1k); arch google/automl

    ResNet   (LibreResNet{18,34,50,101}-cls)           Apache-2.0
             upstream: timm (a1_in1k "ResNet Strikes Back" recipe), ImageNet-1k

    ViT      (LibreViT{ti,s,b,l}-cls)                  Apache-2.0
             upstream: timm AugReg ImageNet-21k pretraining + ImageNet-1k
             fine-tuning; metadata wrap only, learned tensors unchanged

    Swin     (LibreSwin{t,s,b,l}-cls)                  MIT
             upstream: microsoft/Swin-Transformer via timm; t/s/b ImageNet-1k,
             l ImageNet-22k pretraining to ImageNet-1k fine-tuning
             Convert with weights/convert_swin_weights.py.

    CLIP     (LibreCLIP{b32,b16}-cls)                  MIT (weights)
             upstream: mlfoundations/open_clip LAION-2B checkpoints;
             see libreyolo/models/clip/NOTICE.md for the LAION
             data-provenance note.

    SigLIP2  (LibreSigLIP2{b16,so400m}-cls)            Apache-2.0
             upstream: google/siglip2-base-patch16-256 and
             google/siglip2-so400m-patch14-384 (Hugging Face).
             Convert with weights/convert_siglip2_weights.py.

    OMDet-Turbo (LibreOMDetTurbot)                     Apache-2.0
             upstream: omlab/omdet-turbo-swin-tiny-hf
             pinned revision:
               7fe93cecfb770c4d76cf71163956221249cab566
             LibreYOLO mirror revision:
               d569ff1da5668156f07b5859246be435804fed2d
             Model, configuration, processor, and tokenizer assets are
             byte-identical to the pinned upstream snapshot; no learned
             parameters were changed. model.safetensors SHA-256:
               439d1575d7e237ad565ed6969ea2a2dfcedf2086155e9ca3ac96cd6180a48cfd

    OV-DEIM  (LibreOVDEIM{s,m,l})                      CC BY-NC 4.0
             upstream: wleilei/OV-DEIM; converted and rehosted with
             attribution as the upstream MODEL_LICENSE permits;
             licensing confirmed by the author (wleilei/OV-DEIM#4).
             The text tower is apple/MobileCLIP-B-LT-OpenCLIP with
             tensors unmodified; the Apple license text ships in the
             weight repositories.
             Convert with weights/convert_ovdeim_weights.py.

    Depth Anything 3 (LibreDepthAnything3l-depth)      Apache-2.0
             upstream: depth-anything/DA3MONO-LARGE
             pinned revision: f465978e618db8cc79c83b8bbf24964857db1875
             Tensors are unchanged; conversion strips the outer model.
             prefix and adds LibreYOLO checkpoint metadata. Metric and
             any-view checkpoints are not hosted. CC-BY-NC-4.0
             Large/Giant/Nested checkpoints are explicitly excluded.

    MiDaS    (LibreMiDaS{s,l}-depth)                   upstream download only
             upstream: isl-org/MiDaS release assets
             midas_v21_small_256.pt SHA-256:
               70d6b9c891758c67f974a6097fb0c608c7ee67fb81ac3e5588847d5596d56fca
             dpt_large_384.pt SHA-256:
               2f21e586477d90cb9624c7eef5df7891edca49a1c4795ee2cb631fd4daa6ca69
             The releasing repository is MIT, but LibreYOLO does not bundle
             or rehost these checkpoints while commercial-use clearance of
             the mixed training datasets remains unresolved under ADR 0006.
             Runtime auto-download uses the official GitHub release URLs,
             verifies these hashes, safely loads the raw state dict, and adds
             LibreYOLO metadata locally. See docs/provenance/midas.md.

    MoGe-2  (LibreMoGe2{s,b,l}-normal)                 MIT
             upstream / pinned revision / model.pt SHA-256:
             Ruicheng/moge-2-vits-normal
               679230677b4d282c6f304189a93e98e14f085902
               79a16621928c2bf0ed04659218c55c01075e950507f40bb3332fb4c873d3e1dc
             Ruicheng/moge-2-vitb-normal
               54ad3a693e61907ea4633d13dec6ee682fa09419
               16b8110e86d5dc5a849db120ca96ef3a223fd30b0c9146d1d81db504073da5f6
             Ruicheng/moge-2-vitl-normal
               b135031bae30b5ac2ae141a0e68717795ce38340
               280741fd09bc3f403ccff9967784c2a391b52d2c0742ae3efdb21d9f90cc1a01
             All three upstream Hugging Face model cards declare license: mit.
             LibreYOLO downloads the exact pinned objects directly. Conversion
             retains the encoder, shared neck, and normal-head tensors
             unchanged, removes the unused point/mask/metric-scale heads, and
             adds LibreYOLO checkpoint metadata. Convert with
             weights/convert_moge2_weights.py.

    HRNet   (LibreHRNet{w32,w48}-pose)                 MIT implied by release
             upstream: leoxiaobin/deep-high-resolution-net.pytorch at
             6f69e4676ad8d43d0d61b64b1b9726f0c369e7b1 (MIT).
             Official source files:
               pose_hrnet_w32_256x192.pth
                 SHA-256 19bc083708bb8d873211e50d85d56344c10290c6e8b564c813fdde09645c4c1c
               pose_hrnet_w48_384x288.pth
                 SHA-256 95e0fec3194826d5e3f806ea89be68bbb84517b114c3a32b3058c56610b5ef61
             The official model zoo distributes these files from the MIT
             project but attaches no separate per-file license. LibreYOLO
             therefore describes the weight license as implied by the
             releasing project, not checkpoint-specific or independently
             publisher-confirmed. Every mirror includes the verbatim upstream
             MIT LICENSE and this limitation in NOTICE. Conversion preserves
             all learned tensors and adds only checkpoint metadata. See
             docs/provenance/hrnet.md for converted hashes and evidence.

    NAFNet   (LibreNAFNet{s,l}-restore)                no bundled weights
             upstream code: megvii-research/NAFNet (MIT)
             Some published GoPro-trained checkpoint files do not carry an
             explicit standalone weights license. Convert only weights that
             you have the right to use and redistribute. The SIDD denoising
             weights are trained on SIDD (MIT-licensed dataset).

    RealESRGAN (LibreRealESRGAN{x4,x2,x4t}-restore)    no bundled weights
             upstream code: xinntao/Real-ESRGAN (BSD-3-Clause),
             architecture lineage XPixelGroup/BasicSR (Apache-2.0)
             Released generator weights (RealESRGAN_x4plus, RealESRGAN_x2plus,
             realesr-general-x4v3) are BSD-3-Clause; mirrored as converted
             checkpoints with provenance in the model cards.

    SwinIR  (LibreSwinIR{s,m,l}-restore)               Apache-2.0
             upstream: JingyunLiang/SwinIR
             Official SwinIR-S x4 lightweight and SwinIR-M/L x4 real-world
             checkpoints. Convert with weights/convert_swinir_weights.py.

    BiRefNet (LibreBiRefNet{t,l}-matte)                see per-size note
             upstream: ZhengPeng7/BiRefNet (code MIT)
             l (general): Hugging Face weights tagged MIT -> rehosted under
               the LibreYOLO org (LibreYOLO/LibreBiRefNetl-matte).
             t (lite): the BiRefNet_lite HF repo shows an MIT badge in its
               model card but has no explicit license metadata (no YAML
               license field, no LICENSE file). LibreYOLO does NOT rehost the
               lite weights pending an explicit license confirmation; the
               converter and a ready-to-run upload script are provided so the
               maintainer can host them once the tag is confirmed. Users may
               still convert BiRefNet_lite locally at their own discretion.

    FeyNobg (LibreFeyNobgl-matte)                      Apache-2.0
             upstream: feyninc/FeyNobg on Hugging Face (nobg library on
             GitHub), code and weights explicitly Apache-2.0 (YAML license
             tag) -> rehosted under the LibreYOLO org
             (LibreYOLO/LibreFeyNobgl-matte), including
             post-training-quantized fp8/nvfp4 variants
             (LibreFeyNobgl-matte-{fp8,nvfp4}). Upstream training data is
             not disclosed; only the released weights are redistributed.
             Convert with weights/convert_feynobg_weights.py.

    SegFormer (LibreSegformer{b0,b1,b2,b3,b4,b5}-sem)  NON-COMMERCIAL weights
             upstream code: huggingface/transformers modeling_segformer.py
             (Apache-2.0); weights: NVIDIA's ADE20K checkpoints
             (nvidia/segformer-b0..b5-finetuned-ade-*, b5 = 640-640) under
             the NVIDIA Source Code License. That license allows
             redistribution (license copy + attribution must travel with the
             weights) but restricts USE to research or evaluation only, and
             the restriction is viral through derivative works. Hosted with a
             loud non-commercial notice printed before auto-download, like the
             VisDrone research-preview weights. NOT covered by LibreYOLO's
             permissive license. Train from scratch for unrestricted use.

    PP-OCRv5 (LibrePPOCR{t,l}-ocr)                     Apache-2.0
             upstream: PaddlePaddle/PaddleOCR (code and released
             weights Apache-2.0). Converted from the official
             PP-OCRv5_{mobile,server}_{det,rec} training checkpoints;
             exact source files and SHA-256 hashes are recorded on the
             Hugging Face model cards and in
             weights/convert_ppocr_weights.py. The PP-OCRv5 recognition
             dictionary ships inside each checkpoint as charset
             metadata (Apache-2.0, attribution in
             libreyolo/models/ppocr/NOTICE).

    SAM 3    (facebook/sam3 snapshot; not rehosted)   custom SAM License
             upstream: facebook/sam3 on Hugging Face. Access is gated and
             requires accepting Meta's terms. LibreYOLO downloads the weights
             directly and does not distribute them under the MIT license.

    EdgeTAM (LibreYOLO/LibreEdgeTAM)                   Apache-2.0
             Converted from facebook/EdgeTAM revision
             14d7ecc48c656b94e5184519f698cd5386c5a2bf (edgetam.pt
             SHA-256 ed2d4850b8792c239689b043c47046ec239b6e808a3d9b6ae676c803fd8780df).
             The official EdgeTAM repository explicitly licenses both code
             and checkpoints under Apache-2.0. Conversion uses the pinned
             Apache-2.0 Transformers mapping at commit
             bd37c453544e83eb875ed3608980a1660376007a and is verified against
             yonigozlan/EdgeTAM-hf revision
             c266ce53b3fc00f0f495b583f6a116c4e57f53bb.
             Hash-pinned configuration and processor files are copied from
             that Apache-2.0-declared reference; learned tensors are converted
             independently from the official checkpoint and checked exactly.

    PicoSAM3 (LibreYOLO/LibrePicoSAM3)                 Apache-2.0
             Converted and rehosted by LibreYOLO. Upstream sources:
             upstream code commit:
               1b03949e43472953bb0021685c7fc3f5fdf48fde
             upstream weights (pietrobonazzi/picosam3) revision:
               af49e4322b6b7cf448499fee5c073d4576f59444
             Only PicoSAM3_SAM3_student_best.pt matches the advertised
             PicoSAM3 architecture. The epoch-1 files are not mirrored or
             converted because they contain the older PicoSAM2 network.
             Distilled by upstream from SAM 2.1 and SAM 3; teacher weights
             are not included or redistributed.

The LICENSE and NOTICE files on each Hugging Face model card are the
authoritative source for redistribution terms. Users converting
third-party checkpoints locally are responsible for complying with the
upstream license of those checkpoints.

TEED and DexiNed edge specialists
---------------------------------
The TEED and DexiNed architecture source is MIT-licensed. LibreYOLO bundles
the architecture ports and local checkpoint converters, but does not bundle,
mirror, or auto-download the authors' released checkpoints. Those checkpoints
were trained on BIPED, whose published dataset terms restrict use to
non-commercial purposes. Converting a locally held checkpoint does not change
its applicable terms. A checkpoint trained independently on compatibly
licensed data may be distributed under the terms established by its trainer.

Third-party SOURCE CODE bundled inside the LibreYOLO Python package
(separate from weights) is disclosed in the top-level NOTICE file.
Notably, the DEIMv2 family bundles DINOv3 backbone code under Meta's
non-OSI DINOv3 License Agreement; see NOTICE and
libreyolo/models/deimv2/engine/backbone/dinov3/LICENSE.md.
