Roboflow (Official)

Removed value: -{
-  "type": "string"
-}

Added value: +[
+  {
+    "additionalProperties": {
+      "type": "string"
+    },
+    "type": "object"
+  },
+  {
+    "type": "null"
+  }
+]

Added value: +null

Previous value: -"Mapping of class name -> text prompt used to label, e.g. {'cat': 'a cat', 'dog': 'a dog'}. Required. For sam3, max 50 classes."New value: +"Mapping of class name -> text prompt used to label, e.g. {'cat': 'a cat', 'dog': 'a dog'}. Required for foundational models (sam3: bare nouns, max 50 classes). Optional for model_type='roboflow' — when omitted, the trained model's own classes are used."

Removed value: -"object"

Previous value: -[
-  "project_id",
-  "batch_id",
-  "model",
-  "model_type",
-  "ontology"
-]New value: +[
+  "project_id",
+  "batch_id",
+  "model",
+  "model_type"
+]

Previous value: -"Checkpoint to initialize from"New value: +"Checkpoint to initialize from (COCO, a Universe model, or a previous version)"

Previous value: -"Number of training epochs"New value: +"Number of training epochs (max 300)"

Previous value: -"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. For NAS sweeps (one Training produces many child Models): 'rfdetr-nas-parent', 'rfdetr-nas-base-parent', 'rfdetr-nas-seg-parent'. NAS requires the version's validation split to have ≥15 images. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."New value: +"Model architecture ID. Examples: 'rfdetr-medium' (object detection, recommended), 'yolov11n' (fast YOLO), 'rfdetr-seg-medium' (instance segmentation), 'yolo26n-pose' (keypoint), 'vit-base-patch16-224-in21k' (classification), 'qwen3_5-2b-peft' (multimodal/VLM). For NAS sweeps: 'rfdetr-nas-parent' (requires ≥15 validation images). If an invalid ID is passed, the API returns the full list of valid model types for the project's task. Read the training-and-evaluation skill for the complete list and model selection guidance."

Previous value: -"Training speed"New value: +"Training speed (deprecated — model_type is usually sufficient)"

Previous value: -"Checkpoint to initialize from (COCO, a Universe model, or a previous version)"New value: +"Checkpoint to initialize from"

Previous value: -"Number of training epochs (max 300)"New value: +"Number of training epochs"

Previous value: -"Model architecture ID. Examples: 'rfdetr-medium' (object detection, recommended), 'yolov11n' (fast YOLO), 'rfdetr-seg-medium' (instance segmentation), 'yolo26n-pose' (keypoint), 'vit-base-patch16-224-in21k' (classification), 'qwen3_5-2b-peft' (multimodal/VLM). For NAS sweeps: 'rfdetr-nas-parent' (requires ≥15 validation images). If an invalid ID is passed, the API returns the full list of valid model types for the project's task. Read the training-and-evaluation skill for the complete list and model selection guidance."New value: +"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. For NAS sweeps (one Training produces many child Models): 'rfdetr-nas-parent', 'rfdetr-nas-base-parent', 'rfdetr-nas-seg-parent'. NAS requires the version's validation split to have ≥15 images. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."

Previous value: -"Training speed (deprecated — model_type is usually sufficient)"New value: +"Training speed"

Previous value: -"Model architecture ID. Use EXACT IDs below — wrong IDs cause failures.\n\nObject Detection (recommended → RF-DETR):\n  rfdetr-pico, rfdetr-nano, rfdetr-small, rfdetr-base, rfdetr-medium, rfdetr-large, rfdetr-xlarge, rfdetr-2xlarge\n  yolo26n/s/m/l/x, yolov12n/s/m/l/x, yolov11n/s/m/l/x, yolov8n/s/m/l/x\n  yolo_nas_s/m/l, yololite-n/s/m/l/xl, yololite-edge-n/s/m/l/xl\n\nInstance Segmentation:\n  rfdetr-seg-nano/small/medium/large/xlarge/2xlarge\n  yolo26n-seg .. yolo26x-seg, yolov11n-seg .. yolov11x-seg, yolov8n-seg .. yolov8x-seg\n  sam3-large\n\nKeypoint / Pose:\n  yolo26n-pose .. yolo26x-pose, yolov11n-pose .. yolov11x-pose, yolov8n-pose .. yolov8x-pose\n\nClassification:\n  vit-base-patch16-224-in21k, resnet18/34/50/101\n  vit_base_patch16_dinov3.lvd1689m, vit_small_patch16_dinov3.lvd1689m\n\nSemantic Segmentation: deeplabv3plus\n\nMultimodal / VLM:\n  qwen3_5-2b-peft, qwen3_5-0.8b-peft, qwen3vl-2b-instruct, qwen3vl-2b-instruct-peft\n  smolvlm2-peft, smolvlm-256m-peft\n  florence-2-base, florence-2-large, florence-2-base-peft, florence-2-large-peft\n  paligemma2-3b-pt-224/448/896, paligemma2-3b-pt-224-peft\n  qwen25-vl-7b, qwen25-vl-7b-peft\n\nNAS sweeps (produces many child models):\n  rfdetr-nas-parent, rfdetr-nas-base-parent, rfdetr-nas-seg-parent\n  NAS requires ≥15 validation images.\n\nRead the training-and-evaluation skill for model selection guidance, checkpoint options, and training best practices."New value: +"Model architecture ID. Examples: 'rfdetr-medium' (object detection, recommended), 'yolov11n' (fast YOLO), 'rfdetr-seg-medium' (instance segmentation), 'yolo26n-pose' (keypoint), 'vit-base-patch16-224-in21k' (classification), 'qwen3_5-2b-peft' (multimodal/VLM). For NAS sweeps: 'rfdetr-nas-parent' (requires ≥15 validation images). If an invalid ID is passed, the API returns the full list of valid model types for the project's task. Read the training-and-evaluation skill for the complete list and model selection guidance."

Previous value: -"Checkpoint to initialize from"New value: +"Checkpoint to initialize from (COCO, a Universe model, or a previous version)"

Previous value: -"Number of training epochs"New value: +"Number of training epochs (max 300)"

Previous value: -"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. For NAS sweeps (one Training produces many child Models): 'rfdetr-nas-parent', 'rfdetr-nas-base-parent', 'rfdetr-nas-seg-parent'. NAS requires the version's validation split to have ≥15 images. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."New value: +"Model architecture ID. Use EXACT IDs below — wrong IDs cause failures.\n\nObject Detection (recommended → RF-DETR):\n  rfdetr-pico, rfdetr-nano, rfdetr-small, rfdetr-base, rfdetr-medium, rfdetr-large, rfdetr-xlarge, rfdetr-2xlarge\n  yolo26n/s/m/l/x, yolov12n/s/m/l/x, yolov11n/s/m/l/x, yolov8n/s/m/l/x\n  yolo_nas_s/m/l, yololite-n/s/m/l/xl, yololite-edge-n/s/m/l/xl\n\nInstance Segmentation:\n  rfdetr-seg-nano/small/medium/large/xlarge/2xlarge\n  yolo26n-seg .. yolo26x-seg, yolov11n-seg .. yolov11x-seg, yolov8n-seg .. yolov8x-seg\n  sam3-large\n\nKeypoint / Pose:\n  yolo26n-pose .. yolo26x-pose, yolov11n-pose .. yolov11x-pose, yolov8n-pose .. yolov8x-pose\n\nClassification:\n  vit-base-patch16-224-in21k, resnet18/34/50/101\n  vit_base_patch16_dinov3.lvd1689m, vit_small_patch16_dinov3.lvd1689m\n\nSemantic Segmentation: deeplabv3plus\n\nMultimodal / VLM:\n  qwen3_5-2b-peft, qwen3_5-0.8b-peft, qwen3vl-2b-instruct, qwen3vl-2b-instruct-peft\n  smolvlm2-peft, smolvlm-256m-peft\n  florence-2-base, florence-2-large, florence-2-base-peft, florence-2-large-peft\n  paligemma2-3b-pt-224/448/896, paligemma2-3b-pt-224-peft\n  qwen25-vl-7b, qwen25-vl-7b-peft\n\nNAS sweeps (produces many child models):\n  rfdetr-nas-parent, rfdetr-nas-base-parent, rfdetr-nas-seg-parent\n  NAS requires ≥15 validation images.\n\nRead the training-and-evaluation skill for model selection guidance, checkpoint options, and training best practices."

Previous value: -"Training speed"New value: +"Training speed (deprecated — model_type is usually sufficient)"

Previous value: -"NAS-trained model id (Firestore document id, from models_list[].id or trainings_get_results[].models[].modelId)."New value: +"Public model id (the `url` field from models_list[] or models[].modelId from trainings_get_results), e.g. 'beer-can-hackathon-410-nas-gpu-b'. Just the bare id — workspace prefix added automatically."

Added value: +{
+  "anyOf": [
+    {
+      "type": "string"
+    },
+    {
+      "type": "null"
+    }
+  ],
+  "default": null,
+  "description": "Optional NAS modelGroup to scope the list to a single NAS run. Get this value from trainings_get_results."
+}

Previous value: -"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."New value: +"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. For NAS sweeps (one Training produces many child Models): 'rfdetr-nas-parent', 'rfdetr-nas-base-parent', 'rfdetr-nas-seg-parent'. NAS requires the version's validation split to have ≥15 images. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."

Previous value: -"Model architecture (e.g. 'yolov8n', 'rf-detr-base', 'paligemma2-3b')"New value: +"Model architecture. Recommended: RF-DETR ('rfdetr-base', 'rfdetr-nano', 'rfdetr-large', etc). Also: 'yolov11n', 'yolov8n', 'yolo26n', etc. Read the training-and-evaluation skill for the complete list of valid model IDs — wrong IDs cause failures."