Jakub
67061cf373
Document six completed LR runs as preparation for full Aya training on Perun. Record selection of LR 0.00010 by lowest final validation loss, selected-model lm-eval and partial SK-bench results, reproducible provenance, and dataset-access/context-truncation limitations. Preserve existing 30k and Titan material.
3.7 KiB
3.7 KiB
| 1 | model_id | training_samples | model_name | source_result_json | arc_acc_norm | hellaswag_acc_norm | mmlu_acc | truthfulqa_mc1_acc | truthfulqa_mc2_acc | sklegal_acc_norm | skquad_f1 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2 | 01_baseline_r16_a16_lr5e5_adamw | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_01_baseline_r16_a16_lr5e5_adamw | lmeval/mistral_aya(hruska)/01_baseline_r16_a16_lr5e5_adamw/results_2026-09-30T21-23-33.435952.json | 0.34046193327630453 | 0.5209277807063785 | 0.3549992344204563 | 0.17994858611825193 | 0.34391089322085056 | 0.46851574212893554 | 36.6797346829585 |
| 3 | 02_optimizer_adamw8bit_r16_a16_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_02_optimizer_adamw8bit_r16_a16_lr5e5 | lmeval/mistral_aya(hruska)/02_optimizer_adamw8bit_r16_a16_lr5e5/results_2026-10-01T04-23-23.791749.json | 0.3378956372968349 | 0.521244069583553 | 0.35438677078548464 | 0.1928020565552699 | 0.35373417730948353 | 0.4715142428785607 | 36.21782844597198 |
| 4 | 03_rank8_a16_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_03_rank8_a16_lr5e5 | lmeval/mistral_aya(hruska)/03_rank8_a16_lr5e5/results_2026-10-01T10-25-51.531765.json | 0.33105218135158254 | 0.5167105956773853 | 0.34550604807839536 | 0.18380462724935734 | 0.3424520772406517 | 0.4715142428785607 | 36.36378811056779 |
| 5 | 04_rank32_a32_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_04_rank32_a32_lr5e5 | lmeval/mistral_aya(hruska)/04_rank32_a32_lr5e5/results_2026-10-01T17-22-48.146390.json | 0.33532934131736525 | 0.5200843437005799 | 0.3449701423977951 | 0.19023136246786632 | 0.3367013908863517 | 0.47301349325337333 | 36.02282635500946 |
| 6 | 05_r16_alpha32_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_05_r16_alpha32_lr5e5 | lmeval/mistral_aya(hruska)/05_r16_alpha32_lr5e5/results_2026-10-02T00-19-55.725179.json | 0.33361847733105215 | 0.5190300474433316 | 0.35714285714285715 | 0.19665809768637532 | 0.3555125811399826 | 0.46701649175412296 | 37.873456685535025 |
| 7 | 06_lr1e4_r16_a16 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_06_lr1e4_r16_a16 | lmeval/mistral_aya(hruska)/06_lr1e4_r16_a16/results_2026-10-02T06-59-24.237777.json | 0.33105218135158254 | 0.5207169214549289 | 0.3549992344204563 | 0.19023136246786632 | 0.34733281230911506 | 0.46551724137931033 | 37.57962398015711 |
| 8 | 07_lr2e5_r16_a16 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_07_lr2e5_r16_a16 | lmeval/mistral_aya(hruska)/07_lr2e5_r16_a16/results_2026-10-02T14-27-49.730611.json | 0.32763045337895635 | 0.522509225092251 | 0.3379268105956209 | 0.16709511568123395 | 0.32463255972324384 | 0.47226386806596704 | 33.83807391933327 |
| 9 | 08_attention_only_r16_a16_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_08_attention_only_r16_a16_lr5e5 | lmeval/mistral_aya(hruska)/08_attention_only_r16_a16_lr5e5/results_2026-10-02T19-20-02.720498.json | 0.35243798118049613 | 0.5217712177121772 | 0.3448170264890522 | 0.18766066838046272 | 0.3673805917101876 | 0.4767616191904048 | 43.37517960476381 |
| 10 | 09_batch2_accum4_r16_a16_lr5e5 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_09_batch2_accum4_r16_a16_lr5e5 | lmeval/mistral_aya(hruska)/09_batch2_accum4_r16_a16_lr5e5/results_2026-10-03T01-58-48.046423.json | 0.3378956372968349 | 0.5209277807063785 | 0.35140101056499773 | 0.18637532133676094 | 0.340221777957301 | 0.46926536731634183 | 35.3128443275733 |
| 11 | 10_r16_a32_lr1e4 | 100000 | /home/schwarc/diplomovka/hruska_compare9_100k_fixed_10_r16_a32_lr1e4 | lmeval/mistral_aya(hruska)/10_r16_a32_lr1e4/results_2026-10-03T07-58-30.464881.json | 0.32934131736526945 | 0.515656299420137 | 0.3448170264890522 | 0.20565552699228792 | 0.34795284436431645 | 0.4617691154422789 | 37.75612784500817 |
| 12 | aya_300k_lr_0p00010 | 300000 | /home/schwarc/diplomovka/skbench_evaluation/models/aya_300k_lr_0p00010 | aya_300k_lr_experiments/benchmarks/lmeval/results/results_2026-10-24T02-28-32.987157.json | 0.3473053892215569 | 0.5163943068002108 | 0.3102893890675241 | 0.19665809768637532 | 0.3506542164316134 | 0.46101949025487254 | 38.72920807013749 |