NeevVision-v1-cleaned

Prepared from keyur-neeved/NeevVision-v1 at revision 0affe3d628b5ed2449d5423a690030406d98a23a.

Input: 811,652 - kept: 760,907 - exact duplicates removed: 43,409

• near duplicates removed: 7,336 - undecodable images excluded: 0.

class_id	class_name	binary_label	binary_name
0	real	0	real
1	FS	1	fake
2	FR	1	fake
3	EFS	1	fake
4	FE	1	fake
5	PA	1	fake

Images: EXIF-oriented RGB, aspect ratio kept, longer side at most 256 px (Lanczos, never enlarged), lossless PNG. Transparency is flattened onto white. Every row keeps its original split; no re-splitting or balancing.

Dedup (same class only, across all splits, priority test > validation > train): exact = identical RGB pixels; near = 256-bit perceptual-hash shortlist, then same output size and, on a copy with the longer side <= 128 px: RGB SSIM >= 0.95, mean |pixel error| <= 0.02, worst 4x4-tile SSIM >= 0.9, worst tile error <= 0.04. No face identity, alignment, rotation or mirror matching, so different views of a face are kept. A removed row always matches a kept row directly. SSIM 0.95 is a structural score, not "95 % identical pixels"; the shortlist is approximate, so some near duplicates can remain.

split	class	input	kept	exact removed	near removed	errors
test	real	20,192	20,185	0	7	0
test	FS	43,498	43,387	4	107	0
test	FR	32,981	32,770	20	191	0
test	EFS	23,681	22,501	1,177	3	0
validation	real	20,209	20,194	0	15	0
validation	FS	42,416	42,219	10	187	0
validation	FR	33,214	32,673	21	520	0
validation	EFS	23,682	20,297	3,377	8	0
train	real	96,419	96,293	10	116	0
train	FS	206,609	205,562	68	979	0
train	FR	158,237	152,964	148	5,125	0
train	EFS	110,514	71,862	38,574	78	0

reports/ lists every removed row with the kept row it duplicates, image errors, cross-label identical images and these counts.