-
Add curated local racing gear crops
-
Drop 37 ambiguous/badly-cropped samples where model disagrees with label
-
Clean labels using Qwen3-VL-4B VLM verification
-
fix gear 4: replace noisy labels with 300 augmented from clean reference crop
- Removed all noisy sebring/racing-original label=4 (full of 1s and 3s)
- Generated 250 train + 50 val augmented 4s from verified reference crop
- Augmentations: affine skew, random crop, brightness/contrast, noise, blur
- Saved clean reference as examples/sebring-tobi-lap6/gear_4.png
- 5,505 train / 948 val -
v3: deep clean via per-source cross-val + manual curation
- Per-source iterative cross-val cleaning (3 rounds each)
- racing-original: 99.6% OOF, removed all mislabeled 4s (were actually 1s)
- sebring: 82% OOF, removed 91 noisy transition frames
- mnist: 99% OOF, removed 30 ambiguous digits
- Fixed Paul Ricard gear_4 label CSV (was pointing to a 3)
- Fixed gear_4 example image
- Per-label verification composites in composites/{train,validation}/
- 5,391 train / 903 val -
add per-label verification composites and cleaning scripts
- composites/train/label_N.png — racing-only training samples per digit
- composites/validation/label_N.png — racing-only val samples per digit
- scripts/make_composites.py — regenerate composites from parquet
- scripts/iterative_clean.py — multi-round cross-val + ensemble cleaning -
clean: remove 445 noisy train + 99 noisy val samples via ensemble model
Train cleaned with 5-fold cross-val (92.3% OOF), val cleaned with
5-model ensemble trained on clean train (88.6% agreement).
5,368 train / 924 val. -
val: keep only 2 MNIST per digit (340→20), focus val on real racing
-
move 20 more racing 4s into val set (19→39 real racing samples)
-
cleanup: remove temp files, fix find_label_issues.py guard
-
clean: fix 36 mislabels, remove 122 ambiguous samples via cleanlab
- Fixed 36 racing transition frames with wrong labels (trusted model prediction)
- Removed 41 ambiguous MNIST, 7 bad augmented, 74 borderline racing samples
- 5,842 train samples (was 5,964), validation unchanged
- OOF accuracy 89%, cleanlab flagged down from 158 to ~80 issues -
fix: build_dataset.py writes HF Image feature type
-
fix: restore HF Image feature type in parquet schema
-
v2: add new sources, extraction pipeline, better train/val split
- Added Sebring Q Tobi Lap 6 and Paul Ricard Alpine LMPh video sources
- Proper stratified 85/15 train/val split (was nearly empty val before)
- Augmented minority racing classes (gears 4, 7) to 200 samples
- 5,964 train / 1,003 val images (up from 5,321 / 318)
- Added scripts/extract.py and scripts/build_dataset.py for pipeline
- Added scripts/find_label_issues.py (cleanlab-based QA)
- Added examples/ with reference crops per gear per source
- Added labels/ CSVs for frame-range labeling
- Added AGENTS.md with domain context and crop coordinates -
Update README with full dataset documentation
-
Initial dataset: racing gear digits + MNIST supplement
-
initial commit
