alpindale/haruka-compressed

Haruka Tagger, compressed

This is Haruka Tagger (EVA02-B/14 at 448 px, 98,782 tags) compressed from 809 MB to 187 MB for downloading into a browser. foidfinder runs it in the browser with WebGPU, at 52 ms per image on an M4 Pro in Chrome, and with MLX on Apple Silicon (foidfinder --model-dir <local copy> serve).

Accuracy

Agreement with the fp32 model on 400 recent safebooru.org posts, at the recommended thresholds (character, copyright and general 0.35, artist 0.25). The counts are changed decisions.

	size	KL per image	max |Δp|	character	copyright	artist	general	top-1 character
fp16	386 MB	0.0007	0.008	2	1	0	33	100%
bf16, the training precision	386 MB	0.046	0.079	9	5	10	242	99.7%
this checkpoint	187 MB	0.023	0.031	5	6	2	175	100%

KL is summed over all tags.

Method

The backbone linears, the pool's key/value projection and the tag classifier hold 99.6% of the weights. GPTQ, calibrated on 1,536 other safebooru posts, rounds each of them to integer multiples of one fp16 step per output row. The step is the row's RMS divided by 40 in the backbone and by 32 in the classifier. The integer codes are zigzag-mapped, split into byte planes and gzipped. They average 7.4 bits per weight. Everything else is fp16.

foidfinder quantize-haruka KaraKaraWitch/haruka-tagger <dir> --calib <images> reproduces it.

Format

model.safetensors holds the weights in the layout that foidfinder's runtimes read. The input normalization is folded into the patch embedding and the attention scales into the query and key projections. QKV is one matrix, and so are the two SwiGLU input projections. Linear weights are [out, in].

For each compressed matrix name with config.json["quantization"]["layers"][name] = {"shape": [rows, cols], "planes": n}:

• name.step, f16 [rows]
• name.plane0 … name.plane{n-1}, u8, each a gzip stream of rows × cols bytes. Byte i of z is in plane i.

z      = plane0[j] | plane1[j] << 8 | ...
code   = (z >> 1) ^ -(z & 1)
weight = code * step[row]          # row = j / cols

config.json also holds the architecture, the tag names and categories in output order, and the recommended thresholds.