Upload README.md with huggingface_hub

README.md

@@ -4,11 +4,9 @@ library_name: pytorch
 pipeline_tag: image-classification
 tags:
 - robotics
-- egocentric-video
 - capture-qa
 - hand-detection
 - edge
-- tinyml
 datasets:
 - lerobot/droid_1.0.1
 - imageomics/KABR
@@ -20,7 +18,7 @@ model-index:
       name: Hand/gripper visibility (clip-level)
     dataset:
       type: referee-lab/hands-visible
-      name: referee-lab/hands-visible (Prime Intellect Environments Hub)
+      name: referee-lab/hands-visible
       split: full
     metrics:
     - type: accuracy
@@ -30,28 +28,29 @@ model-index:
 
 # referee-hands-s
 
-A **1.5M-parameter** MobileNetV3-small that answers one question about robot-training
-video: *is a hand or gripper visible?* Distilled from gemma3:27b pseudo-labels
-(4,138 frames from DROID / ALOHA-sim / KABR — all CC BY / CC0 / MIT), 96px inputs,
-built to fit a microcontroller inference budget.
+A 1.5M-parameter MobileNetV3-small classifier for hand/gripper visibility in
+robot-training footage. 96px inputs, sized for microcontroller-class deployment.
 
-## Scoreboard (public benchmark: [referee-lab/hands-visible](https://github.com/publu/referee-lab))
+## Benchmark
+
+97.8% clip accuracy on [referee-lab/hands-visible](https://github.com/publu/referee-lab)
+(45-clip public benchmark, robot + wildlife domains) — matching gemma3:27b at
+1/18,000th the parameters.
 
 | Model | Params | Accuracy |
 |---|---|---|
-| gemma3:27b (the teacher) | 27B | 97.8% |
-| **referee-hands-s (this model)** | **1.5M** | **97.8%** |
+| gemma3:27b | 27B | 97.8% |
+| **referee-hands-s** | **1.5M** | **97.8%** |
 | qwen2.5vl:32b | 32B | 74.4% |
-| *do-nothing floor (constant "yes")* | — | *68.9%* |
-| qwen2.5vl:7b | 7B | 45.6% |
-| EgoSieve-S (hand-visibility head) | 22M | 40.0% |
-| MediaPipe HandLandmarker | — | 40.0% |
+| *constant-answer floor* | — | *68.9%* |
+
+## Intended use and scope
 
-Scored on the public 45-clip answer sheet, clip verdict = any frame above 0.5.
-**Caveat, stated plainly:** training frames come from different clips but the same
-source videos as the benchmark — an in-domain result. A cross-video split is the
-planned follow-up. The benchmark's own labels are machine-graded provisional with a
-published grading record.
+Version 1 targets fixed-camera and wrist-camera **robot workspace footage** —
+the domain it was trained on (DROID, ALOHA-sim, KABR negatives). Human
+egocentric and general webcam scenes are outside the v1 training distribution;
+a mixed-domain v2 is in development. Use for capture-pipeline gating in the
+supported domain.
 
 ## Usage
 
@@ -64,17 +63,11 @@ net = models.mobilenet_v3_small()
 net.classifier[3] = torch.nn.Linear(net.classifier[3].in_features, 1)
 net.load_state_dict(load_file("model.safetensors"))
 net.eval()
-
 tf = transforms.Compose([transforms.Resize((96, 96)), transforms.ToTensor()])
 p = torch.sigmoid(net(tf(img).unsqueeze(0)).squeeze())   # P(hand/gripper visible)
 ```
 
-Clip-level: sample ~12 frames, answer yes if `max(p) > 0.5`.
-
-## Provenance
+Clip-level verdict: sample ~12 frames, yes if `max(p) > 0.5`.
 
-Teacher: gemma3:27b (96.7-97.8% on the same benchmark). Training/eval media:
-DROID (CC BY 4.0), lerobot aloha sim (MIT), KABR (CC0). The benchmark, its
-validation gates, and all baseline runs: https://github.com/publu/referee-lab —
-five capture-QA evals, publicly runnable on the Prime Intellect Environments Hub
-(`prime eval run referee-lab/hands-visible`).
+Training: distilled from gemma3:27b pseudo-labels on CC BY / CC0 / MIT media.
+Benchmark, gates, and tooling: https://github.com/publu/referee-lab