Flint
Collection
RecursiveSI's Flint (39.8B, 2.5B active) quantized to GGUF by AtomicChat, with its metrics. • 2 items • Updated • 1
text stringlengths 0 165 |
|---|
moe-hybrid: flint, 49 blocks, mtp [], bands {'edge': [2, 3, 46, 47, 48], 'mid': []} |
label ftype rules overr GiB BPW file_type |
AD-Q8_0 Q8_0 7 0 39.97 8.63 None |
AD-Q6_K Q6_K 7 358 32.34 6.98 None |
AD-Q5_K_L Q5_K_M 7 400 28.68 6.19 None |
AD-Q5_K_M Q5_K_M 7 358 26.98 5.83 None |
AD-Q5_K_S Q5_K_S 7 343 26.49 5.72 None |
AD-Q4_K_L Q4_K_M 7 400 25.42 5.49 None |
AD-Q4_K_M Q4_K_M 7 400 24.03 5.19 None |
AD-IQ4_NL IQ4_NL 7 400 24.03 5.19 None |
AD-IQ4_XS IQ4_XS 7 390 22.82 4.93 None |
AD-Q4_K_S Q4_K_S 7 358 22.55 4.87 None |
AD-IQ4_XXS IQ4_XS 7 348 21.17 4.57 None |
AD-IQ3_M IQ3_M 7 400 19.04 4.11 None |
AD-IQ3_S IQ3_S 7 358 17.78 3.84 None |
AD-IQ3_XXS IQ3_XXS 7 400 17.61 3.80 None |
AD-IQ2_M IQ2_M 7 400 15.23 3.29 None |
AD-IQ2_S IQ2_S 7 484 14.61 3.15 None |
AD-IQ2_XS IQ2_XS 7 400 13.95 3.01 None |
AD-IQ2_XXS IQ2_XXS 7 400 12.37 2.67 None |
AD-IQ1_M IQ1_M 7 400 11.55 2.49 None |
AD-IQ1_S IQ1_S 7 400 10.34 2.23 None |
AD-NVFP4 NVFP4 7 343 22.08 4.77 None |
Q4_K_FLAT Q4_K_S 7 343 22.08 4.77 None |
IQ4_XS_FLAT IQ4_XS 7 343 20.98 4.53 None |
wrote 23 rungs to /work/Flint/ladder |
AD-IQ2_M: ok, 779 tensors as planned, 400 overrides, 15594.16 MiB 3.29 BPW, llama.cpp 19ab5a5e6 |
AD-IQ2_XXS: ok, 779 tensors as planned, 400 overrides, 12664.16 MiB 2.67 BPW, llama.cpp 19ab5a5e6 |
AD-IQ3_S: ok, 779 tensors as planned, 358 overrides, 18206.16 MiB 3.84 BPW, llama.cpp 19ab5a5e6 |
AD-IQ4_XS: ok, 779 tensors as planned, 390 overrides, 23362.16 MiB 4.93 BPW, llama.cpp 19ab5a5e6 |
AD-Q4_K_M: ok, 779 tensors as planned, 400 overrides, 24604.16 MiB 5.19 BPW, llama.cpp 19ab5a5e6 |
AD-Q6_K: ok, 779 tensors as planned, 358 overrides, 33121.08 MiB 6.98 BPW, llama.cpp 19ab5a5e6 |
AD-Q8_0: ok, 779 tensors as planned, 0 overrides, 40933.08 MiB 8.63 BPW, llama.cpp 19ab5a5e6 |
0.01.438.514 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_gate_shexp.weight' |
0.01.447.867 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_up_exps.weight' |
0.01.497.567 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_down_shexp.weight' |
0.01.499.870 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_up_shexp.weight' |
0.01.499.873 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_gate_inp.weight' |
0.01.499.943 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_down_exps.weight' |
0.01.520.514 W compute_tensor_statistics: no preceding-layer tensor for 'blk.2.ffn_gate_exps.weight' |
Computing tensor statistics for /work/Flint/imatrix/imatrix.gguf (531 tensors) |
Layer | Tensor | Mean StdDev Skew Kurt H Norm | ∑ E[A²] Gain | L2 Dist PCC Cov |
--------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
0 | blk.0.attn_k.weight | -0.0209 1.4508 0.1590 0.8578 99.48% | 32653.2407 nan | nan nan nan |
0 | blk.0.attn_output.weight | 0.0045 0.2572 0.2151 1.9730 91.09% | 1095.1840 nan | nan nan nan |
0 | blk.0.attn_q.weight | -0.0209 1.4508 0.1590 0.8578 99.48% | 32653.2407 nan | nan nan nan |
0 | blk.0.attn_v.weight | -0.0209 1.4508 0.1590 0.8578 99.48% | 32653.2407 nan | nan nan nan |
0 | blk.0.ffn_down.weight | -0.0012 0.0933 -0.2605 6.3686 98.89% | 260.2698 nan | nan nan nan |
0 | blk.0.ffn_gate.weight | -0.0518 2.2922 0.4036 4.7418 98.18% | 37910.2650 nan | nan nan nan |
0 | blk.0.ffn_up.weight | -0.0518 2.2922 0.4036 4.7418 98.18% | 37910.2650 nan | nan nan nan |
1 | blk.1.attn_k.weight | -0.0269 0.8610 -0.6292 16.3653 98.37% | 9136.7977 0.53 | 69.6030 0.1924 0.2403 |
1 | blk.1.attn_output.weight | 0.0027 0.0846 -0.1025 0.7681 91.55% | 115.0409 0.32 | 17.1371 0.0374 0.0008 |
1 | blk.1.attn_q.weight | -0.0269 0.8610 -0.6292 16.3653 98.37% | 9136.7977 0.53 | 69.6030 0.1924 0.2403 |
1 | blk.1.attn_v.weight | -0.0269 0.8610 -0.6292 16.3653 98.37% | 9136.7977 0.53 | 69.6030 0.1924 0.2403 |
1 | blk.1.ffn_down.weight | 0.0001 0.0279 0.0138 17.9951 98.48% | 34.2048 0.36 | 7.6559 -0.0115 -0.0000 |
1 | blk.1.ffn_gate.weight | -0.0349 1.4968 1.4099 25.3269 97.48% | 20908.5504 0.74 | 96.1691 0.4341 1.4894 |
1 | blk.1.ffn_up.weight | -0.0349 1.4968 1.4099 25.3269 97.48% | 20908.5504 0.74 | 96.1691 0.4341 1.4894 |
2 | blk.2.attn_k.weight | -0.0270 0.6646 -1.0869 12.7055 98.08% | 4247.9696 0.68 | 33.0347 0.5681 0.3250 |
2 | blk.2.attn_output.weight | -0.0010 0.0315 -0.1502 2.3800 93.31% | 16.5340 0.38 | 5.7425 0.0214 0.0001 |
2 | blk.2.attn_q.weight | -0.0270 0.6646 -1.0869 12.7055 98.08% | 4247.9696 0.68 | 33.0347 0.5681 0.3250 |
2 | blk.2.attn_v.weight | -0.0270 0.6646 -1.0869 12.7055 98.08% | 4247.9696 0.68 | 33.0347 0.5681 0.3250 |
2 | blk.2.ffn_down_exps.weight | 0.0000 0.0102 0.9516 240.9892 87.10% | 250.4399 nan | nan nan nan |
2 | blk.2.ffn_down_shexp.weight | 0.0006 0.0172 -0.1239 0.6605 99.45% | 2.5934 nan | nan nan nan |
2 | blk.2.ffn_gate_exps.weight | -0.0078 1.5052 -0.0494 4.9186 98.90% | 4087685.6524 nan | nan nan nan |
2 | blk.2.ffn_gate_inp.weight | -0.0086 1.3379 -0.0582 6.1204 98.63% | 15710.8554 nan | nan nan nan |
2 | blk.2.ffn_gate_shexp.weight | -0.0086 1.3379 -0.0582 6.1204 98.63% | 15710.8554 nan | nan nan nan |
2 | blk.2.ffn_up_exps.weight | -0.0078 1.5052 -0.0494 4.9186 98.90% | 4087685.6524 nan | nan nan nan |
2 | blk.2.ffn_up_shexp.weight | -0.0086 1.3379 -0.0582 6.1204 98.63% | 15710.8554 nan | nan nan nan |
3 | blk.3.attn_k.weight | 0.0023 0.5828 -0.3784 3.2760 99.22% | 3743.4533 0.94 | 18.6536 0.7904 0.3061 |
3 | blk.3.attn_output.weight | 0.0008 0.0179 0.0539 1.7950 95.89% | 9.5542 0.76 | 2.2506 0.0722 0.0000 |
3 | blk.3.attn_q.weight | 0.0023 0.5828 -0.3784 3.2760 99.22% | 3743.4533 0.94 | 18.6536 0.7904 0.3061 |
3 | blk.3.attn_v.weight | 0.0023 0.5828 -0.3784 3.2760 99.22% | 3743.4533 0.94 | 18.6536 0.7904 0.3061 |
3 | blk.3.ffn_down_exps.weight | -0.0000 0.0078 1.5717 162.8231 92.36% | 274.9908 1.05 | 4.6508 -0.0010 -0.0000 |
3 | blk.3.ffn_down_shexp.weight | -0.0002 0.0129 -0.2181 2.9546 99.26% | 1.7192 0.81 | 0.6855 0.0074 0.0000 |
3 | blk.3.ffn_gate_exps.weight | 0.0215 1.7000 -0.0089 0.1845 99.55% | 6906675.9963 1.30 | 1010.8058 0.6268 1.6039 |
3 | blk.3.ffn_gate_inp.weight | 0.0193 1.5282 -0.0056 0.2187 99.60% | 26417.9335 1.30 | 43.1170 0.7871 1.6093 |
3 | blk.3.ffn_gate_shexp.weight | 0.0193 1.5282 -0.0056 0.2187 99.60% | 26417.9335 1.30 | 43.1170 0.7871 1.6093 |
3 | blk.3.ffn_up_exps.weight | 0.0215 1.7000 -0.0089 0.1845 99.55% | 6906675.9963 1.30 | 1010.8058 0.6268 1.6039 |
3 | blk.3.ffn_up_shexp.weight | 0.0193 1.5282 -0.0056 0.2187 99.60% | 26417.9335 1.30 | 43.1170 0.7871 1.6093 |
4 | blk.4.attn_k.weight | 0.0077 0.5509 -1.0140 10.9281 99.09% | 3814.6768 1.01 | 13.2733 0.8677 0.2785 |
4 | blk.4.attn_output.weight | 0.0006 0.0133 0.3455 1.5992 96.68% | 4.9075 0.72 | 1.3754 0.0782 0.0000 |
4 | blk.4.attn_q.weight | 0.0077 0.5509 -1.0140 10.9281 99.09% | 3814.6768 1.01 | 13.2733 0.8677 0.2785 |
4 | blk.4.attn_v.weight | 0.0077 0.5509 -1.0140 10.9281 99.09% | 3814.6768 1.01 | 13.2733 0.8677 0.2785 |
4 | blk.4.ffn_down_exps.weight | -0.0000 0.0077 -4.8436 322.1180 95.07% | 309.1652 1.06 | 3.9735 0.0036 0.0000 |
4 | blk.4.ffn_down_shexp.weight | -0.0003 0.0093 -0.6309 6.8611 97.12% | 0.9908 0.76 | 0.5145 -0.0216 -0.0000 |
4 | blk.4.ffn_gate_exps.weight | 0.0367 1.7797 -0.0249 0.1634 99.61% | 8272391.2340 1.09 | 927.5591 0.7299 2.2083 |
4 | blk.4.ffn_gate_inp.weight | 0.0336 1.6152 -0.0369 0.2210 99.67% | 31824.5878 1.10 | 32.6307 0.8963 2.2124 |
4 | blk.4.ffn_gate_shexp.weight | 0.0336 1.6152 -0.0369 0.2210 99.67% | 31824.5878 1.10 | 32.6307 0.8963 2.2124 |
4 | blk.4.ffn_up_exps.weight | 0.0367 1.7797 -0.0249 0.1634 99.61% | 8272391.2340 1.09 | 927.5591 0.7299 2.2083 |
4 | blk.4.ffn_up_shexp.weight | 0.0336 1.6152 -0.0369 0.2210 99.67% | 31824.5878 1.10 | 32.6307 0.8963 2.2124 |
5 | blk.5.attn_k.weight | 0.0077 0.4704 -0.5897 5.6288 99.37% | 2876.4120 0.87 | 9.4109 0.9290 0.2408 |
5 | blk.5.attn_output.weight | 0.0000 0.0071 0.0535 2.0688 94.83% | 1.3346 0.52 | 0.9653 0.0036 0.0000 |
5 | blk.5.attn_q.weight | 0.0077 0.4704 -0.5897 5.6288 99.37% | 2876.4120 0.87 | 9.4109 0.9290 0.2408 |
5 | blk.5.attn_v.weight | 0.0077 0.4704 -0.5897 5.6288 99.37% | 2876.4120 0.87 | 9.4109 0.9290 0.2408 |
5 | blk.5.ffn_down_exps.weight | 0.0000 0.0093 1.3508 436.5443 93.42% | 425.7468 1.17 | 4.3916 -0.0046 -0.0000 |
5 | blk.5.ffn_down_shexp.weight | 0.0004 0.0112 0.2363 2.8798 97.47% | 1.1389 1.07 | 0.4711 -0.0182 -0.0000 |
5 | blk.5.ffn_gate_exps.weight | 0.0433 1.8011 -0.0677 0.4529 99.60% | 9074064.8800 1.05 | 835.5057 0.7924 2.5399 |
End of preview. Expand in Data Studio
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, onneutralandcode;results/ext/: the same for other publishers' files, measured against the same reference;kld/base-<set>.kld.*.partandkld/base-<set>.manifest.txt: the reference logits of the BF16 weights (join the parts withcat) 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.
- Downloads last month
- -