text stringlengths 0 1.44k |
|---|
[2026-06-14 06:50:05] |
============================================================ |
[2026-06-14 06:50:05] GGUF MODEL EVALUATION |
[2026-06-14 06:50:05] ============================================================ |
[2026-06-14 06:50:05] === CONFIG START === |
[2026-06-14 06:50:05] base_repo: "unsloth/Qwen3.5-4B-GGUF" |
[2026-06-14 06:50:05] base_file: "Qwen3.5-4B-BF16.gguf" |
[2026-06-14 06:50:05] ctx_size: 512 |
[2026-06-14 06:50:05] batch_size: 512 |
[2026-06-14 06:50:05] n_parallel: 1 |
[2026-06-14 06:50:05] seed: 1337 |
[2026-06-14 06:50:05] chunks: 50 |
[2026-06-14 06:50:05] dataset_repo: null |
[2026-06-14 06:50:05] dataset_split: null |
[2026-06-14 06:50:05] test_files: [] |
[2026-06-14 06:50:05] target_models: [["unsloth/Qwen3.5-4B-GGUF", "Qwen3.5-4B-UD-IQ2_M.gguf"], ["unsloth/Qwen3.5-4B-GGUF", "Qwen3.5-4B-UD-IQ2_XXS.gguf"], ["unsloth/Qwen3.5-4B-GGUF", "Qwen3.5-4B-UD-IQ3_XXS.gguf"], ["unsloth/Qwen3.5-4B-GGUF", "Qwen3.5-4B-UD-Q2_K_XL.gguf"], ["unsloth/Qwen3.5-4B-GGUF", "Qwen3.5-4B-UD-Q3_K_... |
[2026-06-14 06:50:05] status: "NEW" |
[2026-06-14 06:50:05] === CONFIG END === |
[2026-06-14 06:50:05] Eval dir : models/evaluation |
[2026-06-14 06:50:05] Temp dir : /kaggle/working/ |
[2026-06-14 06:50:05] DAT dir : /tmp/llama_dat |
[2026-06-14 06:50:05] Binary dir: /workspace/llama_bins/cuda-12.8 |
[2026-06-14 06:50:05] Binary : /workspace/llama_bins/cuda-12.8/llama-perplexity |
[2026-06-14 06:50:05] Target models to evaluate: 26 |
[2026-06-14 06:50:05] Using default test file: /tmp/wiki_test.raw |
[2026-06-14 06:50:05] Test files: 1 |
[2026-06-14 06:50:05] |
📥 Downloading base model: unsloth/Qwen3.5-4B-GGUF / Qwen3.5-4B-BF16.gguf |
[2026-06-14 06:50:27] ✅ Base model downloaded: /kaggle/working/Qwen3.5-4B-BF16.gguf |
[2026-06-14 06:50:27] |
============================================================ |
[2026-06-14 06:50:27] PHASE 1: Generating base logits for all files |
[2026-06-14 06:50:27] ============================================================ |
[2026-06-14 06:50:27] [1/1] Base model on /tmp/wiki_test.raw → /tmp/llama_dat/base_logits_0.dat |
[2026-06-14 06:50:27] Running: /workspace/llama_bins/cuda-12.8/llama-perplexity -m /kaggle/working/Qwen3.5-4B-BF16.gguf -f /tmp/wiki_test.raw -c 512 -b 512 -np 1 --kl-divergence-base /tmp/llama_dat/base_logits_0.dat --seed 1337 -ngl 99 --chunks 50 |
[2026-06-14 06:50:28] 0.00.287.191 I common_init_result: fitting params to device memory ... |
[2026-06-14 06:50:28] 0.00.287.215 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on) |
[2026-06-14 06:50:37] 0.09.503.874 W llama_context: n_ctx_seq (512) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
[2026-06-14 06:50:38] 0.09.544.870 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
[2026-06-14 06:51:24] 0.56.403.178 I |
[2026-06-14 06:51:25] 0.56.403.342 I system_info: n_threads = 2 (n_threads_batch = 2) / 4 | CUDA : ARCHS = 750,800,860,890,900,1000,1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | |
[2026-06-14 06:51:25] 0.56.403.441 I perplexity: saving all logits to /tmp/llama_dat/base_logits_0.dat |
[2026-06-14 06:51:25] 0.56.403.448 I perplexity: tokenizing the input .. |
[2026-06-14 06:51:25] 0.57.115.319 I perplexity: tokenization took 711.857 ms |
[2026-06-14 06:51:25] 0.57.115.754 I perplexity: calculating perplexity over 50 chunks, n_ctx=512, batch_size=512, n_seq=1 |
[2026-06-14 06:51:40] 1.11.724.760 I perplexity: 14.45 seconds per pass - ETA 12.03 minutes |
[2026-06-14 06:53:16] [1]5.9954,[2]8.8643,[3]9.0343,[4]8.8617,[5]8.9319,[6]9.2448,[7]9.5683,[8]10.4290,[9]11.0487,[10]11.4813,[11]11.3769,[12]11.5143,[13]12.1970,[14]11.6112,[15]11.4999,[16]11.6183,[17]10.8196,[18]10.8113,[19]10.7109,[20]10.7439,[21]10.3575,[22]10.2251,[23]9.8303,[24]9.3419,[25]9.1660,[26]8.9160,[27]8.... |
[2026-06-14 06:53:17] 2.48.535.419 I Final estimate: PPL = 9.3905 +/- 0.22827 |
[2026-06-14 06:53:17] |
[2026-06-14 06:53:18] |
Phase 1 complete. Valid files: 1 / 1 |
[2026-06-14 06:53:19] 🗑️ Deleted /kaggle/working/Qwen3.5-4B-BF16.gguf to free disk space |
[2026-06-14 06:53:19] ============================================================ |
[2026-06-14 06:53:19] PHASE 2: Running KL divergence for each target model |
[2026-06-14 06:53:19] ============================================================ |
[2026-06-14 06:53:19] |
============================================================ |
[2026-06-14 06:53:19] TARGET MODEL 1/26: unsloth/Qwen3.5-4B-GGUF / Qwen3.5-4B-UD-IQ2_M.gguf |
[2026-06-14 06:53:19] ============================================================ |
[2026-06-14 06:53:28] |
[1/1] Qwen3.5-4B-UD-IQ2_M.gguf vs /tmp/wiki_test.raw (base: /tmp/llama_dat/base_logits_0.dat) |
[2026-06-14 06:53:28] Running: /workspace/llama_bins/cuda-12.8/llama-perplexity -m /kaggle/working/Qwen3.5-4B-UD-IQ2_M.gguf -f /tmp/wiki_test.raw -c 512 -b 512 -np 1 --kl-divergence-base /tmp/llama_dat/base_logits_0.dat --kl-divergence --seed 1337 -ngl 99 --chunks 50 |
[2026-06-14 06:53:28] 0.00.266.961 I common_init_result: fitting params to device memory ... |
[2026-06-14 06:53:28] 0.00.266.969 I common_init_result: (for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on) |
[2026-06-14 06:53:35] 0.06.903.670 W llama_context: n_ctx_seq (512) < n_ctx_train (262144) -- the full capacity of the model will not be utilized |
[2026-06-14 06:53:35] 0.06.941.754 I common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable) |
[2026-06-14 06:53:48] 0.19.596.194 I |
[2026-06-14 06:53:48] 0.19.596.308 I system_info: n_threads = 2 (n_threads_batch = 2) / 4 | CUDA : ARCHS = 750,800,860,890,900,1000,1200 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 | |
[2026-06-14 06:53:48] 0.19.678.281 I kl_divergence: computing over 50 chunks, n_ctx=512, batch_size=512, n_seq=1 |
[2026-06-14 06:54:58] 1.30.144.022 I kl_divergence: 70.47 seconds per pass - ETA 58.72 minutes |
[2026-06-14 06:54:59] |
[2026-06-14 06:54:59] chunk PPL ln(PPL(Q)/PPL(base)) KL Divergence Δp RMS Same top p |
[2026-06-14 06:54:59] 1 8.5392 ± 1.6222 0.35366 ± 0.06274 0.24591 ± 0.03174 17.267 ± 1.876 % 81.961 ± 2.413 % |
[2026-06-14 06:54:59] 2 11.6935 ± 1.6322 0.27721 ± 0.04444 0.26594 ± 0.02587 15.024 ± 1.261 % 80.980 ± 1.740 % |
[2026-06-14 06:55:00] 3 11.8070 ± 1.2932 0.26781 ± 0.03892 0.28770 ± 0.02117 15.546 ± 1.069 % 80.131 ± 1.444 % |
[2026-06-14 06:55:01] 4 11.1840 ± 1.0132 0.23285 ± 0.03089 0.26035 ± 0.01629 14.670 ± 0.904 % 79.804 ± 1.258 % |
[2026-06-14 06:55:01] 5 10.8398 ± 0.8562 0.19368 ± 0.02587 0.24173 ± 0.01326 13.854 ± 0.777 % 79.529 ± 1.130 % |
[2026-06-14 06:55:02] 6 11.2407 ± 0.8000 0.19554 ± 0.02285 0.23619 ± 0.01126 13.252 ± 0.685 % 78.301 ± 1.054 % |
[2026-06-14 06:55:02] 7 11.5064 ± 0.7553 0.18449 ± 0.02044 0.22976 ± 0.00980 12.709 ± 0.617 % 78.599 ± 0.971 % |
[2026-06-14 06:55:03] 8 12.2933 ± 0.7656 0.16450 ± 0.01880 0.22885 ± 0.00906 12.294 ± 0.561 % 78.431 ± 0.911 % |
[2026-06-14 06:55:04] 9 12.9860 ± 0.7666 0.16160 ± 0.01748 0.22373 ± 0.00826 12.017 ± 0.522 % 78.519 ± 0.857 % |
[2026-06-14 06:55:04] 10 13.5504 ± 0.7619 0.16573 ± 0.01636 0.22166 ± 0.00768 11.958 ± 0.484 % 78.235 ± 0.817 % |
[2026-06-14 06:55:05] 11 13.3482 ± 0.7113 0.15982 ± 0.01561 0.22230 ± 0.00742 12.181 ± 0.479 % 78.075 ± 0.781 % |
[2026-06-14 06:55:06] 12 13.5419 ± 0.6920 0.16222 ± 0.01497 0.22743 ± 0.00707 12.444 ± 0.461 % 77.974 ± 0.749 % |
[2026-06-14 06:55:06] 13 14.1313 ± 0.7003 0.14722 ± 0.01458 0.22904 ± 0.00671 12.394 ± 0.442 % 77.526 ± 0.725 % |
[2026-06-14 06:55:07] 14 13.4919 ± 0.6400 0.15058 ± 0.01411 0.23867 ± 0.00750 12.964 ± 0.435 % 77.647 ± 0.697 % |
[2026-06-14 06:55:07] 15 13.3448 ± 0.6119 0.14922 ± 0.01432 0.25507 ± 0.01046 13.492 ± 0.450 % 77.516 ± 0.675 % |
[2026-06-14 06:55:08] 16 13.6168 ± 0.6094 0.15912 ± 0.01425 0.25995 ± 0.01021 13.638 ± 0.440 % 77.279 ± 0.656 % |
[2026-06-14 06:55:09] 17 12.6959 ± 0.5464 0.16030 ± 0.01362 0.25613 ± 0.00967 13.779 ± 0.423 % 77.509 ± 0.634 % |
[2026-06-14 06:55:09] 18 12.7401 ± 0.5313 0.16452 ± 0.01337 0.25875 ± 0.00933 14.044 ± 0.418 % 77.516 ± 0.616 % |
[2026-06-14 06:55:10] 19 12.6743 ± 0.5163 0.16865 ± 0.01310 0.26299 ± 0.00924 14.157 ± 0.411 % 77.730 ± 0.598 % |
[2026-06-14 06:55:11] 20 12.7154 ± 0.5048 0.16879 ± 0.01289 0.26581 ± 0.00899 14.177 ± 0.396 % 77.765 ± 0.582 % |
[2026-06-14 06:55:11] 21 12.2609 ± 0.4724 0.16901 ± 0.01250 0.26433 ± 0.00862 14.203 ± 0.383 % 77.908 ± 0.567 % |
[2026-06-14 06:55:12] 22 12.1005 ± 0.4540 0.16869 ± 0.01216 0.26470 ± 0.00827 14.297 ± 0.371 % 77.897 ± 0.554 % |
[2026-06-14 06:55:13] 23 11.6594 ± 0.4252 0.17092 ± 0.01176 0.26159 ± 0.00797 14.353 ± 0.360 % 78.090 ± 0.540 % |
[2026-06-14 06:55:13] 24 11.0657 ± 0.3912 0.16962 ± 0.01139 0.25926 ± 0.00771 14.630 ± 0.358 % 78.317 ± 0.527 % |
[2026-06-14 06:55:14] 25 10.8554 ± 0.3739 0.16941 ± 0.01111 0.25897 ± 0.00747 14.838 ± 0.351 % 78.447 ± 0.515 % |
[2026-06-14 06:55:15] 26 10.5460 ± 0.3536 0.16815 ± 0.01084 0.25794 ± 0.00724 14.809 ± 0.341 % 78.492 ± 0.505 % |
[2026-06-14 06:55:15] 27 10.3563 ± 0.3393 0.16895 ± 0.01063 0.25722 ± 0.00702 14.802 ± 0.332 % 78.577 ± 0.495 % |
End of preview. Expand in Data Studio
No dataset card yet
- Downloads last month
- 48