AtomicChat/Flint-GGUF-metrics

Flint GGUF metrics

Everything behind AtomicChat/Flint-GGUF: the files there are built with our own importance matrix, and every number on its card comes from here.

• results.json: size, KL divergence, top-1 and perplexity of every rung, on neutral and code;
• results/ext/: the same for other publishers' files, measured against the same reference;
• kld/base-<set>.kld.*.part and kld/base-<set>.manifest.txt: the reference logits of the BF16 weights (join the parts with cat) and the protocol they were written with;
• bf16/: the BF16 GGUF conversion the rungs were quantized from;
• imatrix/: the importance matrix, its statistics, its parameters and the corpus manifest;
• ladder/: the tensor-type rules of every rung; logs/: quantization, imatrix and KL-divergence logs;
• inventory.json: the tensors of the BF16 conversion.

fast/ is the first set of files, published on 22 October with bartowski's importance matrix and replaced the same night: their rules, quantization logs and a quick perplexity check.