Datasets:
summary dict | results list |
|---|---|
{
"created_at": "2026-06-06T17:16:04",
"input_dir": "/home/moon/Downloads/Test-100",
"repo_id": "Seoyoung07/korean-sign-word-classifier-mediapipe",
"revision": "main",
"snapshot": "/home/moon/.cache/huggingface/hub/models--Seoyoung07--korean-sign-word-classifier-mediapipe/snapshots/f7f31077b55c4aa6e64ba58d2db... | [
{
"index": 1,
"file": "/home/moon/Downloads/Test-100/간호사.mp4",
"expected_raw": "간호사",
"expected": "간호사",
"label_alias_applied": false,
"expected_in_vocab": true,
"predicted": "간호사",
"confidence": 0.9996203184127808,
"top_k": [
{
"rank": 1,
"label": "간호사",
... |
KSL Test-100 Evaluation Dataset
This dataset contains 100 isolated Korean sign-language word videos used for the Test-100 evaluation of the MediaPipe keypoint word classifier.
Contents
videos/: MP4 isolated Korean sign-language word clips.metadata.csv: Video-level labels, model predictions, and evaluation fields.evaluation/: Full evaluation reports and machine-readable result files.
Evaluation Summary
- Model:
Seoyoung07/korean-sign-word-classifier-mediapipe - Model snapshot used for evaluation:
214c0d87294695f4c79086440404bfb72fa7c5bc - Checkpoint epoch:
41 - Checkpoint validation accuracy:
0.8962213765938981 - Device used for evaluation:
cuda - Total videos: 100
- Top-1 correct: 60 / 100 (60.00%)
- Top-5 contains expected: 71 / 100 (71.00%)
- Inference errors: 0
AI Tool Used
- AI tool used: Codex
Notes
The filename typo 꺠끗하다.mp4 is kept as the original filename, while metadata.csv records the canonical label as 깨끗하다. Evaluation reports also document this normalization.
License And Data Use
This dataset is published for the realtime KSL captioning project evaluation workflow. Confirm any institutional, privacy, or presentation requirements before redistributing or reusing the videos outside that context.
- Downloads last month
- 135