Maggio33 commited on
Commit
a9841ba
·
verified ·
1 Parent(s): b070442

Slayer1B 1.09: update cards and documentation

Browse files
Files changed (4) hide show
  1. DATA_CARD.md +3 -1
  2. EVALUATION.json +509 -1
  3. README.md +23 -23
  4. TRAINING_CONFIGURATION.md +3 -3
DATA_CARD.md CHANGED
@@ -18,7 +18,9 @@ tags:
18
 
19
  ## Target, units and exposure
20
 
21
- **Funding and planning update — 11 November 2026, 15:29–15:41 UTC:** an additional **USD 500 was reported received** for continued main-base training. This is not the current account balance; the earlier USD 425 envelope is a dated planning record, not a new remaining-budget calculation. The goal remains maximum useful affordable exposure with constant learning rates after warmup and **no in-run terminal decay**, retaining cost guards and the operational reserve. The approved operator recipe selects `KROKI=120000` with decay from update 96,000; at the 15:41 UTC snapshot the running process had not yet adopted it and still used 80,000. Track's separate 80,000-update plan is an estimate, not a funded endpoint or evidence of execution. Review the schedule before its decay boundary becomes reachable; do not bring decay forward by shortening the scheduling argument. Source allocation targets 20% PL / 80% EN, not the cumulative model-language ratio.
 
 
22
 
23
  **Prepared-data inventory — 11 November 2026, 14:32 UTC, v6 token positions:** the dataset operator reported the following aggregate counts. Prepared input capacity, reported training progress and unfinished material are different states; rows overlap as labelled and must not all be summed.
24
 
 
18
 
19
  ## Target, units and exposure
20
 
21
+ **Funding and runtime update — 11 November 2026:** an additional **USD 500 was reported received at 15:29 UTC** for continued main-base training. This is not the current account balance; the earlier USD 425 envelope is historical, not a new remaining-budget calculation. The goal remains maximum useful affordable exposure with constant learning rates after warmup and **no in-run terminal decay**, retaining cost guards and the operational reserve. Source allocation targets 20% PL / 80% EN, not the cumulative model-language ratio.
22
+
23
+ The operator recipe was installed on the workstation at 15:41 UTC; **runtime adoption was independently confirmed after restart at 16:23:47 UTC**, with `RESUME @ 29800`, `--steps 120000`, full checkpoints every 200 updates and learning rate 0.0006. The configured decay boundary is update 96,000. This dated receipt confirms adoption, not present hardware availability or a funded endpoint. Track's separate 80,000-update plan remains a revisable estimate. Review the schedule before its decay boundary becomes reachable; do not bring decay forward by shortening the scheduling argument.
24
 
25
  **Prepared-data inventory — 11 November 2026, 14:32 UTC, v6 token positions:** the dataset operator reported the following aggregate counts. Prepared input capacity, reported training progress and unfinished material are different states; rows overlap as labelled and must not all be summed.
26
 
EVALUATION.json CHANGED
@@ -2,7 +2,7 @@
2
  "schema_version": 1,
3
  "measurement_owner": "SlayerLab / Fabryka AI",
4
  "status": "descriptive; training-lineage contamination scan pending",
5
- "source_aggregate_sha256": "fc6f67a20bdc6ada687349d35de26f6811e27766dfe1aa197082000c03535796",
6
  "source_generator_sha256": null,
7
  "benchmark_community_report": "https://github.com/slayerlabs/benchmarks-fabryka-ai/blob/8b547007526971bd2fc35dfe56ff2c7d9b802568/research/results/full-run-2026-11/README.md",
8
  "community_report_is_separate_measurement": true,
@@ -1573,6 +1573,514 @@
1573
  "correct_primary": 3351
1574
  }
1575
  }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1576
  }
1577
  },
1578
  "additional_measurements": {
 
2
  "schema_version": 1,
3
  "measurement_owner": "SlayerLab / Fabryka AI",
4
  "status": "descriptive; training-lineage contamination scan pending",
5
+ "source_aggregate_sha256": "2d257a95523605b06aa964cb974f65d86ab025875cc1f87dc98b242b90729440",
6
  "source_generator_sha256": null,
7
  "benchmark_community_report": "https://github.com/slayerlabs/benchmarks-fabryka-ai/blob/8b547007526971bd2fc35dfe56ff2c7d9b802568/research/results/full-run-2026-11/README.md",
8
  "community_report_is_separate_measurement": true,
 
1573
  "correct_primary": 3351
1574
  }
1575
  }
1576
+ },
1577
+ "1.09-k25000": {
1578
+ "step": 25000,
1579
+ "weights_published": false,
1580
+ "weights_sha256": "46dfedda0d855f36f86fb1a0c0661809c654422c76bd1b3b2259e081f115b310",
1581
+ "evaluator_sha256": "ae6658f9c83a484e373dc6d14476a7aeb98a905c27cd6d97ea234fb56140e016",
1582
+ "source_results_sha256": "7b1ac4f7710b11b64542eb193a2820144c3ba6482901159b20356d2928e6f9c5",
1583
+ "tasks": {
1584
+ "arc_challenge": {
1585
+ "n": 1172,
1586
+ "n_pelne": 1172,
1587
+ "dataset_repo": "allenai/ai2_arc",
1588
+ "dataset_sha": "210d026faf9955653af8916fad021475a3f00453",
1589
+ "pliki_sha256": {
1590
+ "ARC-Challenge/test-00000-of-00001.parquet": "62f03257e737aed263f55c6abf87c7bb0028a44a6bdd2a26eb1279eb42c1d1e9"
1591
+ },
1592
+ "acc": 0.313993,
1593
+ "acc_ci95": [
1594
+ 0.288396,
1595
+ 0.340444
1596
+ ],
1597
+ "acc_norm": 0.337884,
1598
+ "acc_norm_ci95": [
1599
+ 0.31058,
1600
+ 0.365188
1601
+ ],
1602
+ "split": "test",
1603
+ "correct_primary": 396
1604
+ },
1605
+ "hellaswag": {
1606
+ "n": 10042,
1607
+ "n_pelne": 10042,
1608
+ "dataset_repo": "Rowan/hellaswag",
1609
+ "dataset_sha": "218ec52e09a7e7462a5400043bb9a69a41d06b76",
1610
+ "pliki_sha256": {
1611
+ "data/validation-00000-of-00001.parquet": "899813071e1e95efafec90f856e1987d2150fa4d020fc005df6962c259f660cd"
1612
+ },
1613
+ "acc": 0.397132,
1614
+ "acc_ci95": [
1615
+ 0.387373,
1616
+ 0.406592
1617
+ ],
1618
+ "acc_norm": 0.511651,
1619
+ "acc_norm_ci95": [
1620
+ 0.501792,
1621
+ 0.52141
1622
+ ],
1623
+ "split": "validation",
1624
+ "correct_primary": 5138
1625
+ },
1626
+ "piqa": {
1627
+ "n": 1838,
1628
+ "n_pelne": 1838,
1629
+ "dataset_repo": "baber/piqa",
1630
+ "dataset_sha": "142f6d7367fd9877f0fb3b5734ea6a545f54cdd1",
1631
+ "pliki_sha256": {
1632
+ "piqa_validation.parquet": "2b17ee1215b1ceb2ddabf4c90a05c5a8b833bef9d207960f57d111c43a2080f3"
1633
+ },
1634
+ "acc": 0.706202,
1635
+ "acc_ci95": [
1636
+ 0.685528,
1637
+ 0.726877
1638
+ ],
1639
+ "acc_norm": 0.715452,
1640
+ "acc_norm_ci95": [
1641
+ 0.694763,
1642
+ 0.736126
1643
+ ],
1644
+ "split": "validation",
1645
+ "correct_primary": 1315
1646
+ },
1647
+ "winogrande": {
1648
+ "n": 1267,
1649
+ "n_pelne": 1267,
1650
+ "dataset_repo": "allenai/winogrande",
1651
+ "dataset_sha": "01e74176c63542e6b0bcb004dcdea22d94fb67b5",
1652
+ "pliki_sha256": {
1653
+ "winogrande_xl/validation-00000-of-00001.parquet": "f9d914d1818c0aba98cabdf8af2bc4bd943c462b9c5e7062647e05c2ce3315d1"
1654
+ },
1655
+ "acc": 0.5588,
1656
+ "acc_ci95": [
1657
+ 0.531176,
1658
+ 0.586425
1659
+ ],
1660
+ "acc_norm": 0.5588,
1661
+ "acc_norm_ci95": [
1662
+ 0.531965,
1663
+ 0.586425
1664
+ ],
1665
+ "split": "validation",
1666
+ "correct_primary": 708
1667
+ },
1668
+ "boolq": {
1669
+ "n": 3270,
1670
+ "n_pelne": 3270,
1671
+ "dataset_repo": "aps/super_glue",
1672
+ "dataset_sha": "3de24cf8022e94f4ee4b9d55a6f539891524d646",
1673
+ "pliki_sha256": {
1674
+ "boolq/validation-00000-of-00001.parquet": "f46438b8cce048184552a2bfcdb8252eadb3a2602f0c22334caccd9b6ea547d6"
1675
+ },
1676
+ "acc": 0.522018,
1677
+ "acc_ci95": [
1678
+ 0.505191,
1679
+ 0.539144
1680
+ ],
1681
+ "acc_norm": 0.614679,
1682
+ "acc_norm_ci95": [
1683
+ 0.598165,
1684
+ 0.631498
1685
+ ],
1686
+ "split": "validation",
1687
+ "correct_primary": 1707
1688
+ },
1689
+ "openbookqa": {
1690
+ "n": 500,
1691
+ "n_pelne": 500,
1692
+ "dataset_repo": "allenai/openbookqa",
1693
+ "dataset_sha": "388097ea7776314e93a529163e0fea805b8a6454",
1694
+ "pliki_sha256": {
1695
+ "main/test-00000-of-00001.parquet": "cd5483e366daa230c1c87bbdc512d8b7229f14f6dd04d19fc8b1a3855aaaa8a3"
1696
+ },
1697
+ "acc": 0.26,
1698
+ "acc_ci95": [
1699
+ 0.22,
1700
+ 0.3
1701
+ ],
1702
+ "acc_norm": 0.372,
1703
+ "acc_norm_ci95": [
1704
+ 0.328,
1705
+ 0.414
1706
+ ],
1707
+ "split": "test",
1708
+ "correct_primary": 186
1709
+ },
1710
+ "belebele_pol": {
1711
+ "n": 900,
1712
+ "n_pelne": 900,
1713
+ "dataset_repo": "facebook/belebele",
1714
+ "dataset_sha": "7899cdfa4e1e0d733fd77c848e2c273cb1d32be2",
1715
+ "pliki_sha256": {
1716
+ "data/pol_Latn.jsonl": "1699bb60b65e258d71989c541027c5439eb36a4f5032a5be392137bf5a1b6bec"
1717
+ },
1718
+ "acc": 0.253333,
1719
+ "acc_ci95": [
1720
+ 0.224444,
1721
+ 0.282222
1722
+ ],
1723
+ "acc_norm": 0.253333,
1724
+ "acc_norm_ci95": [
1725
+ 0.224444,
1726
+ 0.282222
1727
+ ],
1728
+ "split": "pol_Latn evaluation set",
1729
+ "correct_primary": 228
1730
+ },
1731
+ "hellaswag_pl": {
1732
+ "n": 10042,
1733
+ "n_pelne": 10042,
1734
+ "dataset_repo": "Eurolingua/hellaswagx",
1735
+ "dataset_sha": "c38911a204759866dbcde35eeb584c36caf11a8b",
1736
+ "pliki_sha256": {
1737
+ "hellaswag_PL_validation.jsonl": "acdfc6d397885a4c86de178a6e5f9b7e9b90a7ff81ffcd6e5e0d3d0788dd3a16"
1738
+ },
1739
+ "acc": 0.342561,
1740
+ "acc_ci95": [
1741
+ 0.333499,
1742
+ 0.351524
1743
+ ],
1744
+ "acc_norm": 0.439355,
1745
+ "acc_norm_ci95": [
1746
+ 0.429397,
1747
+ 0.449114
1748
+ ],
1749
+ "split": "validation",
1750
+ "correct_primary": 4412
1751
+ },
1752
+ "mmlu_pl": {
1753
+ "n": 14042,
1754
+ "n_pelne": 14042,
1755
+ "dataset_repo": "Eurolingua/mmlux",
1756
+ "dataset_sha": "7fb62cf05c1ce9ea550dbfd2ea6c145204344882",
1757
+ "pliki_sha256": {
1758
+ "hendrycks_abstract_algebra_PL_test.jsonl": "4905c6ef8d306c2ff1e1771ef8c15b63152bb88effec11844217893b81fcc3ea",
1759
+ "hendrycks_anatomy_PL_test.jsonl": "5cba9bc54a6f0bdaddcd2e0e23155fccfeab469b561f03cc571e1387abeb9a38",
1760
+ "hendrycks_astronomy_PL_test.jsonl": "00dc5fca7a8be3d8bef9e24ad9018e5417fb07703ebd2d0039c2c82418272b03",
1761
+ "hendrycks_business_ethics_PL_test.jsonl": "9c9542880c720e2d0f5a768ab2d6602bdf5c394d66e51ff352af79aaca6bbd96",
1762
+ "hendrycks_clinical_knowledge_PL_test.jsonl": "79fd5fd7a17dc7c7d84a9e9b138cf13c2dcaf6ab332eb1a0f299a7b52dc89c61",
1763
+ "hendrycks_college_biology_PL_test.jsonl": "b49d766fb2c4160809d3907b846cd6c5c609299f47d9096bcf3208b843392c48",
1764
+ "hendrycks_college_chemistry_PL_test.jsonl": "36c6fd391411789310dca0bbd36aec9aae8784df2bac2e6838e0895eebd97f1d",
1765
+ "hendrycks_college_computer_science_PL_test.jsonl": "90ab29151dc4b8c7e13893516474201b73e8f9765cdf87225b6b9cacdd540e70",
1766
+ "hendrycks_college_mathematics_PL_test.jsonl": "84411ee6464d01f24f2482bccb451635fe75d8169e33932ce6e067c1be3d37ee",
1767
+ "hendrycks_college_medicine_PL_test.jsonl": "faa451d891c8a67b15bae4921c49a8afea448ac7b1b6095646c3570ac3040606",
1768
+ "hendrycks_college_physics_PL_test.jsonl": "bfcaac4cb6d13e7cbcc6ebfb7ea6a9ab0d0ec07613b1b10ca679e632106b4bc2",
1769
+ "hendrycks_computer_security_PL_test.jsonl": "c8324fb94b5b4b7e17440520ea20a22d8c03c3b86da81824059602982463cf8a",
1770
+ "hendrycks_conceptual_physics_PL_test.jsonl": "5227e303a5437f437a4ad9d7eb75917d36051ce635af9f43a92d9525918eac64",
1771
+ "hendrycks_econometrics_PL_test.jsonl": "9bf1d2a2c9eb4f5692b3876919dbc4e679d43d3ea2b64dc8fa2c17703a3e696b",
1772
+ "hendrycks_electrical_engineering_PL_test.jsonl": "a431be1d342cc1eca6aa408688b78751c74621056a4fa4c31e7dfda443e4be6f",
1773
+ "hendrycks_elementary_mathematics_PL_test.jsonl": "9cdb8692fc08d6d8ac5dec24d89f386863b6a02480eb3b8c5be641f15700e667",
1774
+ "hendrycks_formal_logic_PL_test.jsonl": "0f30d48346a0cd2f136463b283d8bd3fe085de67da2a34d90cac689740f9b5be",
1775
+ "hendrycks_global_facts_PL_test.jsonl": "21c753017550f0d31d7c2cdbe82637c877682a9edf8ab913de720e79c82b374f",
1776
+ "hendrycks_high_school_biology_PL_test.jsonl": "7b086933dd97cfc392fba5f78c3523ddc0302cf2b7caed7428a75bfdf9fbb7f2",
1777
+ "hendrycks_high_school_chemistry_PL_test.jsonl": "85f836040875313d60cf9241bd9e827a3cdb770cf159ddf3f7032edb293871ec",
1778
+ "hendrycks_high_school_computer_science_PL_test.jsonl": "034b7b497ada5cc1146c6c5fa02291094d5da512bc0c618ae2b967bd941a70c1",
1779
+ "hendrycks_high_school_european_history_PL_test.jsonl": "023ce8f773cb5dccd86658165a86a5bc1b0d83d53cede5a22c8f7a027582e2a4",
1780
+ "hendrycks_high_school_geography_PL_test.jsonl": "3cce3d6defc5cad9742f9db4407ec5176e6b5d1634b0f8ecb4cfe9f4cabfbe5a",
1781
+ "hendrycks_high_school_government_and_politics_PL_test.jsonl": "9c73c16602cd663acf3b08cdecef6aa8a2c019ed647f02cb970f078d88e3ec7d",
1782
+ "hendrycks_high_school_macroeconomics_PL_test.jsonl": "9425c91f59bd35355262f0c60b04b479a2e700064751efddadb9a28cbc83974d",
1783
+ "hendrycks_high_school_mathematics_PL_test.jsonl": "80c73fa2b882104047260829a57cdca588d1ec9ba690b8a8a5cb87f65283e735",
1784
+ "hendrycks_high_school_microeconomics_PL_test.jsonl": "aebc508914a9b32f5dd1bdc3ebd1375a619791214cb67968ccc6d5712675cbe8",
1785
+ "hendrycks_high_school_physics_PL_test.jsonl": "bba7751511cada0b7b0b07133f2a961e5f046e174e201628a6722a279ab4e47f",
1786
+ "hendrycks_high_school_psychology_PL_test.jsonl": "adf5943b42493ae59f9310028a45adffaace2359d52b8cbb1fa9cc41d6371e45",
1787
+ "hendrycks_high_school_statistics_PL_test.jsonl": "14ffdec0e35d7437d7b2243c3bcbf606843df9ac4eaa044cae95faa3f6e41fdf",
1788
+ "hendrycks_high_school_us_history_PL_test.jsonl": "d8a7946dbf1860b1e99ff0d1948ede18500cc06984ef68786ffed6a8160731f7",
1789
+ "hendrycks_high_school_world_history_PL_test.jsonl": "ae943acd33cf4b23fd3d70d4cc5b0d16f073574c547de91bfc157d552bf63ec0",
1790
+ "hendrycks_human_aging_PL_test.jsonl": "21df1fd8d194a6e453e71e1583626fc723f1c9b44cc571ef7d6074450754ed82",
1791
+ "hendrycks_human_sexuality_PL_test.jsonl": "d7c467fcee3671069986948b69e6ec69c2630173207025ce4925719bb1a880d4",
1792
+ "hendrycks_international_law_PL_test.jsonl": "f0b6f86e54390b63fda0ecf07c7c55f87d30864f6a698c9819f8d8e6eca4d746",
1793
+ "hendrycks_jurisprudence_PL_test.jsonl": "9d7f42f41510da5d835d8df834c4bf6bcd2e03e658cc906ec334810b30ebbe96",
1794
+ "hendrycks_logical_fallacies_PL_test.jsonl": "8b7cd34e9126adcfbc4b9b7bee8345d97352ec892221a0e98fa268210b04ee51",
1795
+ "hendrycks_machine_learning_PL_test.jsonl": "f654c60a51877a9d269471ab01b54b0d4d20f9cc07f7fb57512fb3e90d5a8608",
1796
+ "hendrycks_management_PL_test.jsonl": "628374b8f9beabefe43c2b1dda2c07760eb455d2e15f9e58f58506d61f0fc3bb",
1797
+ "hendrycks_marketing_PL_test.jsonl": "31dc1eef1527c98642080c0b3866ea113bda77c2a03c5155aafe382206ffacdb",
1798
+ "hendrycks_medical_genetics_PL_test.jsonl": "f05fcc33d029460a7f7916ec83debfc746ced5ea931f7a56b63fe1245819e4a4",
1799
+ "hendrycks_miscellaneous_PL_test.jsonl": "94be7854c05aef25ebca4ae1dc562f37172b5346a14fb760b8e89eefea588809",
1800
+ "hendrycks_moral_disputes_PL_test.jsonl": "245ad1cf635274e5ac81559542cb21638b7c729e8a811f76271f4301937d6127",
1801
+ "hendrycks_moral_scenarios_PL_test.jsonl": "d60b08bb71451594385db5ff6ffdef4ca11fec801dc86d681e68f1727a690f9b",
1802
+ "hendrycks_nutrition_PL_test.jsonl": "3cfe618d8e23e3c8e5c752544f07a71b90b4c84d4bc5d673026e8be4de4bf285",
1803
+ "hendrycks_philosophy_PL_test.jsonl": "a1939f9773085bdae7e127fce099afc4153ec4cbe5f66e126f12599feaf76639",
1804
+ "hendrycks_prehistory_PL_test.jsonl": "42eaf871b859f22321dfdec21b6b6b4a3c8f1435c2d4cd2da46dc9fb5ceced88",
1805
+ "hendrycks_professional_accounting_PL_test.jsonl": "1fb6440616a195b94d3de760af5cacd05656667dfac595e2601504cda407c25e",
1806
+ "hendrycks_professional_law_PL_test.jsonl": "fb86b175ecdfe8b56d4991bf6d6705a58f1d89a3efda2a8eb5f01f8082b21e57",
1807
+ "hendrycks_professional_medicine_PL_test.jsonl": "ec818eb13a66e631ab82a10bae1536377dcd4bcadf2321c964b7718c60992080",
1808
+ "hendrycks_professional_psychology_PL_test.jsonl": "72ddab1aaddc26266f4964941334a84d9243f63c329d2d9349bdcaa20efa6bc3",
1809
+ "hendrycks_public_relations_PL_test.jsonl": "2e16835d88ba1ebb6f70a526bd85b9830cb1d3e14e2358a43ce15a3d548232e1",
1810
+ "hendrycks_security_studies_PL_test.jsonl": "c476a634f4f48ff025869158057de726a8080bee253d88921d90252204aa3e4c",
1811
+ "hendrycks_sociology_PL_test.jsonl": "dd56860da95ed793943c9c201f9aa8f5aa8d83753b1bbcbc440fff7ac30f881a",
1812
+ "hendrycks_us_foreign_policy_PL_test.jsonl": "95183b3fcd02d17f5bb7d359b709b3fde13bf6f0f0f1375ab676b0e7dda36786",
1813
+ "hendrycks_virology_PL_test.jsonl": "99e2c6f80907315317619f8c92f8d501ae484619c1169e36d9ad8fa59abfbda7",
1814
+ "hendrycks_world_religions_PL_test.jsonl": "4631dbc58d379cbc46751515f36af682f8bf53db36f3345dbe090d5a00b7b24a"
1815
+ },
1816
+ "acc": 0.242629,
1817
+ "acc_ci95": [
1818
+ 0.235721,
1819
+ 0.249608
1820
+ ],
1821
+ "acc_norm": 0.242629,
1822
+ "acc_norm_ci95": [
1823
+ 0.235579,
1824
+ 0.24968
1825
+ ],
1826
+ "split": "test; micro-average over subjects",
1827
+ "correct_primary": 3407
1828
+ }
1829
+ }
1830
+ },
1831
+ "1.09-k30000": {
1832
+ "step": 30000,
1833
+ "weights_published": false,
1834
+ "weights_sha256": "8a80bb721f6e98677fedcc4ea850896b3f3e5fe158b9f30915e7558bfa1c9764",
1835
+ "evaluator_sha256": "ae6658f9c83a484e373dc6d14476a7aeb98a905c27cd6d97ea234fb56140e016",
1836
+ "source_results_sha256": "724d1fa367dcc57d07d4f6959a1cec40b98ddc6350179fc55248586fa1da6a3e",
1837
+ "tasks": {
1838
+ "arc_challenge": {
1839
+ "n": 1172,
1840
+ "n_pelne": 1172,
1841
+ "dataset_repo": "allenai/ai2_arc",
1842
+ "dataset_sha": "210d026faf9955653af8916fad021475a3f00453",
1843
+ "pliki_sha256": {
1844
+ "ARC-Challenge/test-00000-of-00001.parquet": "62f03257e737aed263f55c6abf87c7bb0028a44a6bdd2a26eb1279eb42c1d1e9"
1845
+ },
1846
+ "acc": 0.312287,
1847
+ "acc_ci95": [
1848
+ 0.285836,
1849
+ 0.33959
1850
+ ],
1851
+ "acc_norm": 0.330205,
1852
+ "acc_norm_ci95": [
1853
+ 0.302901,
1854
+ 0.356655
1855
+ ],
1856
+ "split": "test",
1857
+ "correct_primary": 387
1858
+ },
1859
+ "hellaswag": {
1860
+ "n": 10042,
1861
+ "n_pelne": 10042,
1862
+ "dataset_repo": "Rowan/hellaswag",
1863
+ "dataset_sha": "218ec52e09a7e7462a5400043bb9a69a41d06b76",
1864
+ "pliki_sha256": {
1865
+ "data/validation-00000-of-00001.parquet": "899813071e1e95efafec90f856e1987d2150fa4d020fc005df6962c259f660cd"
1866
+ },
1867
+ "acc": 0.40709,
1868
+ "acc_ci95": [
1869
+ 0.397431,
1870
+ 0.41665
1871
+ ],
1872
+ "acc_norm": 0.520016,
1873
+ "acc_norm_ci95": [
1874
+ 0.510157,
1875
+ 0.529775
1876
+ ],
1877
+ "split": "validation",
1878
+ "correct_primary": 5222
1879
+ },
1880
+ "piqa": {
1881
+ "n": 1838,
1882
+ "n_pelne": 1838,
1883
+ "dataset_repo": "baber/piqa",
1884
+ "dataset_sha": "142f6d7367fd9877f0fb3b5734ea6a545f54cdd1",
1885
+ "pliki_sha256": {
1886
+ "piqa_validation.parquet": "2b17ee1215b1ceb2ddabf4c90a05c5a8b833bef9d207960f57d111c43a2080f3"
1887
+ },
1888
+ "acc": 0.707835,
1889
+ "acc_ci95": [
1890
+ 0.68716,
1891
+ 0.728523
1892
+ ],
1893
+ "acc_norm": 0.706746,
1894
+ "acc_norm_ci95": [
1895
+ 0.686072,
1896
+ 0.727421
1897
+ ],
1898
+ "split": "validation",
1899
+ "correct_primary": 1299
1900
+ },
1901
+ "winogrande": {
1902
+ "n": 1267,
1903
+ "n_pelne": 1267,
1904
+ "dataset_repo": "allenai/winogrande",
1905
+ "dataset_sha": "01e74176c63542e6b0bcb004dcdea22d94fb67b5",
1906
+ "pliki_sha256": {
1907
+ "winogrande_xl/validation-00000-of-00001.parquet": "f9d914d1818c0aba98cabdf8af2bc4bd943c462b9c5e7062647e05c2ce3315d1"
1908
+ },
1909
+ "acc": 0.571429,
1910
+ "acc_ci95": [
1911
+ 0.544574,
1912
+ 0.598264
1913
+ ],
1914
+ "acc_norm": 0.571429,
1915
+ "acc_norm_ci95": [
1916
+ 0.543804,
1917
+ 0.599053
1918
+ ],
1919
+ "split": "validation",
1920
+ "correct_primary": 724
1921
+ },
1922
+ "boolq": {
1923
+ "n": 3270,
1924
+ "n_pelne": 3270,
1925
+ "dataset_repo": "aps/super_glue",
1926
+ "dataset_sha": "3de24cf8022e94f4ee4b9d55a6f539891524d646",
1927
+ "pliki_sha256": {
1928
+ "boolq/validation-00000-of-00001.parquet": "f46438b8cce048184552a2bfcdb8252eadb3a2602f0c22334caccd9b6ea547d6"
1929
+ },
1930
+ "acc": 0.606728,
1931
+ "acc_ci95": [
1932
+ 0.589908,
1933
+ 0.623547
1934
+ ],
1935
+ "acc_norm": 0.622324,
1936
+ "acc_norm_ci95": [
1937
+ 0.605505,
1938
+ 0.638838
1939
+ ],
1940
+ "split": "validation",
1941
+ "correct_primary": 1984
1942
+ },
1943
+ "openbookqa": {
1944
+ "n": 500,
1945
+ "n_pelne": 500,
1946
+ "dataset_repo": "allenai/openbookqa",
1947
+ "dataset_sha": "388097ea7776314e93a529163e0fea805b8a6454",
1948
+ "pliki_sha256": {
1949
+ "main/test-00000-of-00001.parquet": "cd5483e366daa230c1c87bbdc512d8b7229f14f6dd04d19fc8b1a3855aaaa8a3"
1950
+ },
1951
+ "acc": 0.242,
1952
+ "acc_ci95": [
1953
+ 0.204,
1954
+ 0.28
1955
+ ],
1956
+ "acc_norm": 0.362,
1957
+ "acc_norm_ci95": [
1958
+ 0.32,
1959
+ 0.404
1960
+ ],
1961
+ "split": "test",
1962
+ "correct_primary": 181
1963
+ },
1964
+ "belebele_pol": {
1965
+ "n": 900,
1966
+ "n_pelne": 900,
1967
+ "dataset_repo": "facebook/belebele",
1968
+ "dataset_sha": "7899cdfa4e1e0d733fd77c848e2c273cb1d32be2",
1969
+ "pliki_sha256": {
1970
+ "data/pol_Latn.jsonl": "1699bb60b65e258d71989c541027c5439eb36a4f5032a5be392137bf5a1b6bec"
1971
+ },
1972
+ "acc": 0.244444,
1973
+ "acc_ci95": [
1974
+ 0.216667,
1975
+ 0.273333
1976
+ ],
1977
+ "acc_norm": 0.244444,
1978
+ "acc_norm_ci95": [
1979
+ 0.216667,
1980
+ 0.272222
1981
+ ],
1982
+ "split": "pol_Latn evaluation set",
1983
+ "correct_primary": 220
1984
+ },
1985
+ "hellaswag_pl": {
1986
+ "n": 10042,
1987
+ "n_pelne": 10042,
1988
+ "dataset_repo": "Eurolingua/hellaswagx",
1989
+ "dataset_sha": "c38911a204759866dbcde35eeb584c36caf11a8b",
1990
+ "pliki_sha256": {
1991
+ "hellaswag_PL_validation.jsonl": "acdfc6d397885a4c86de178a6e5f9b7e9b90a7ff81ffcd6e5e0d3d0788dd3a16"
1992
+ },
1993
+ "acc": 0.360287,
1994
+ "acc_ci95": [
1995
+ 0.350926,
1996
+ 0.369647
1997
+ ],
1998
+ "acc_norm": 0.457479,
1999
+ "acc_norm_ci95": [
2000
+ 0.447717,
2001
+ 0.467337
2002
+ ],
2003
+ "split": "validation",
2004
+ "correct_primary": 4594
2005
+ },
2006
+ "mmlu_pl": {
2007
+ "n": 14042,
2008
+ "n_pelne": 14042,
2009
+ "dataset_repo": "Eurolingua/mmlux",
2010
+ "dataset_sha": "7fb62cf05c1ce9ea550dbfd2ea6c145204344882",
2011
+ "pliki_sha256": {
2012
+ "hendrycks_abstract_algebra_PL_test.jsonl": "4905c6ef8d306c2ff1e1771ef8c15b63152bb88effec11844217893b81fcc3ea",
2013
+ "hendrycks_anatomy_PL_test.jsonl": "5cba9bc54a6f0bdaddcd2e0e23155fccfeab469b561f03cc571e1387abeb9a38",
2014
+ "hendrycks_astronomy_PL_test.jsonl": "00dc5fca7a8be3d8bef9e24ad9018e5417fb07703ebd2d0039c2c82418272b03",
2015
+ "hendrycks_business_ethics_PL_test.jsonl": "9c9542880c720e2d0f5a768ab2d6602bdf5c394d66e51ff352af79aaca6bbd96",
2016
+ "hendrycks_clinical_knowledge_PL_test.jsonl": "79fd5fd7a17dc7c7d84a9e9b138cf13c2dcaf6ab332eb1a0f299a7b52dc89c61",
2017
+ "hendrycks_college_biology_PL_test.jsonl": "b49d766fb2c4160809d3907b846cd6c5c609299f47d9096bcf3208b843392c48",
2018
+ "hendrycks_college_chemistry_PL_test.jsonl": "36c6fd391411789310dca0bbd36aec9aae8784df2bac2e6838e0895eebd97f1d",
2019
+ "hendrycks_college_computer_science_PL_test.jsonl": "90ab29151dc4b8c7e13893516474201b73e8f9765cdf87225b6b9cacdd540e70",
2020
+ "hendrycks_college_mathematics_PL_test.jsonl": "84411ee6464d01f24f2482bccb451635fe75d8169e33932ce6e067c1be3d37ee",
2021
+ "hendrycks_college_medicine_PL_test.jsonl": "faa451d891c8a67b15bae4921c49a8afea448ac7b1b6095646c3570ac3040606",
2022
+ "hendrycks_college_physics_PL_test.jsonl": "bfcaac4cb6d13e7cbcc6ebfb7ea6a9ab0d0ec07613b1b10ca679e632106b4bc2",
2023
+ "hendrycks_computer_security_PL_test.jsonl": "c8324fb94b5b4b7e17440520ea20a22d8c03c3b86da81824059602982463cf8a",
2024
+ "hendrycks_conceptual_physics_PL_test.jsonl": "5227e303a5437f437a4ad9d7eb75917d36051ce635af9f43a92d9525918eac64",
2025
+ "hendrycks_econometrics_PL_test.jsonl": "9bf1d2a2c9eb4f5692b3876919dbc4e679d43d3ea2b64dc8fa2c17703a3e696b",
2026
+ "hendrycks_electrical_engineering_PL_test.jsonl": "a431be1d342cc1eca6aa408688b78751c74621056a4fa4c31e7dfda443e4be6f",
2027
+ "hendrycks_elementary_mathematics_PL_test.jsonl": "9cdb8692fc08d6d8ac5dec24d89f386863b6a02480eb3b8c5be641f15700e667",
2028
+ "hendrycks_formal_logic_PL_test.jsonl": "0f30d48346a0cd2f136463b283d8bd3fe085de67da2a34d90cac689740f9b5be",
2029
+ "hendrycks_global_facts_PL_test.jsonl": "21c753017550f0d31d7c2cdbe82637c877682a9edf8ab913de720e79c82b374f",
2030
+ "hendrycks_high_school_biology_PL_test.jsonl": "7b086933dd97cfc392fba5f78c3523ddc0302cf2b7caed7428a75bfdf9fbb7f2",
2031
+ "hendrycks_high_school_chemistry_PL_test.jsonl": "85f836040875313d60cf9241bd9e827a3cdb770cf159ddf3f7032edb293871ec",
2032
+ "hendrycks_high_school_computer_science_PL_test.jsonl": "034b7b497ada5cc1146c6c5fa02291094d5da512bc0c618ae2b967bd941a70c1",
2033
+ "hendrycks_high_school_european_history_PL_test.jsonl": "023ce8f773cb5dccd86658165a86a5bc1b0d83d53cede5a22c8f7a027582e2a4",
2034
+ "hendrycks_high_school_geography_PL_test.jsonl": "3cce3d6defc5cad9742f9db4407ec5176e6b5d1634b0f8ecb4cfe9f4cabfbe5a",
2035
+ "hendrycks_high_school_government_and_politics_PL_test.jsonl": "9c73c16602cd663acf3b08cdecef6aa8a2c019ed647f02cb970f078d88e3ec7d",
2036
+ "hendrycks_high_school_macroeconomics_PL_test.jsonl": "9425c91f59bd35355262f0c60b04b479a2e700064751efddadb9a28cbc83974d",
2037
+ "hendrycks_high_school_mathematics_PL_test.jsonl": "80c73fa2b882104047260829a57cdca588d1ec9ba690b8a8a5cb87f65283e735",
2038
+ "hendrycks_high_school_microeconomics_PL_test.jsonl": "aebc508914a9b32f5dd1bdc3ebd1375a619791214cb67968ccc6d5712675cbe8",
2039
+ "hendrycks_high_school_physics_PL_test.jsonl": "bba7751511cada0b7b0b07133f2a961e5f046e174e201628a6722a279ab4e47f",
2040
+ "hendrycks_high_school_psychology_PL_test.jsonl": "adf5943b42493ae59f9310028a45adffaace2359d52b8cbb1fa9cc41d6371e45",
2041
+ "hendrycks_high_school_statistics_PL_test.jsonl": "14ffdec0e35d7437d7b2243c3bcbf606843df9ac4eaa044cae95faa3f6e41fdf",
2042
+ "hendrycks_high_school_us_history_PL_test.jsonl": "d8a7946dbf1860b1e99ff0d1948ede18500cc06984ef68786ffed6a8160731f7",
2043
+ "hendrycks_high_school_world_history_PL_test.jsonl": "ae943acd33cf4b23fd3d70d4cc5b0d16f073574c547de91bfc157d552bf63ec0",
2044
+ "hendrycks_human_aging_PL_test.jsonl": "21df1fd8d194a6e453e71e1583626fc723f1c9b44cc571ef7d6074450754ed82",
2045
+ "hendrycks_human_sexuality_PL_test.jsonl": "d7c467fcee3671069986948b69e6ec69c2630173207025ce4925719bb1a880d4",
2046
+ "hendrycks_international_law_PL_test.jsonl": "f0b6f86e54390b63fda0ecf07c7c55f87d30864f6a698c9819f8d8e6eca4d746",
2047
+ "hendrycks_jurisprudence_PL_test.jsonl": "9d7f42f41510da5d835d8df834c4bf6bcd2e03e658cc906ec334810b30ebbe96",
2048
+ "hendrycks_logical_fallacies_PL_test.jsonl": "8b7cd34e9126adcfbc4b9b7bee8345d97352ec892221a0e98fa268210b04ee51",
2049
+ "hendrycks_machine_learning_PL_test.jsonl": "f654c60a51877a9d269471ab01b54b0d4d20f9cc07f7fb57512fb3e90d5a8608",
2050
+ "hendrycks_management_PL_test.jsonl": "628374b8f9beabefe43c2b1dda2c07760eb455d2e15f9e58f58506d61f0fc3bb",
2051
+ "hendrycks_marketing_PL_test.jsonl": "31dc1eef1527c98642080c0b3866ea113bda77c2a03c5155aafe382206ffacdb",
2052
+ "hendrycks_medical_genetics_PL_test.jsonl": "f05fcc33d029460a7f7916ec83debfc746ced5ea931f7a56b63fe1245819e4a4",
2053
+ "hendrycks_miscellaneous_PL_test.jsonl": "94be7854c05aef25ebca4ae1dc562f37172b5346a14fb760b8e89eefea588809",
2054
+ "hendrycks_moral_disputes_PL_test.jsonl": "245ad1cf635274e5ac81559542cb21638b7c729e8a811f76271f4301937d6127",
2055
+ "hendrycks_moral_scenarios_PL_test.jsonl": "d60b08bb71451594385db5ff6ffdef4ca11fec801dc86d681e68f1727a690f9b",
2056
+ "hendrycks_nutrition_PL_test.jsonl": "3cfe618d8e23e3c8e5c752544f07a71b90b4c84d4bc5d673026e8be4de4bf285",
2057
+ "hendrycks_philosophy_PL_test.jsonl": "a1939f9773085bdae7e127fce099afc4153ec4cbe5f66e126f12599feaf76639",
2058
+ "hendrycks_prehistory_PL_test.jsonl": "42eaf871b859f22321dfdec21b6b6b4a3c8f1435c2d4cd2da46dc9fb5ceced88",
2059
+ "hendrycks_professional_accounting_PL_test.jsonl": "1fb6440616a195b94d3de760af5cacd05656667dfac595e2601504cda407c25e",
2060
+ "hendrycks_professional_law_PL_test.jsonl": "fb86b175ecdfe8b56d4991bf6d6705a58f1d89a3efda2a8eb5f01f8082b21e57",
2061
+ "hendrycks_professional_medicine_PL_test.jsonl": "ec818eb13a66e631ab82a10bae1536377dcd4bcadf2321c964b7718c60992080",
2062
+ "hendrycks_professional_psychology_PL_test.jsonl": "72ddab1aaddc26266f4964941334a84d9243f63c329d2d9349bdcaa20efa6bc3",
2063
+ "hendrycks_public_relations_PL_test.jsonl": "2e16835d88ba1ebb6f70a526bd85b9830cb1d3e14e2358a43ce15a3d548232e1",
2064
+ "hendrycks_security_studies_PL_test.jsonl": "c476a634f4f48ff025869158057de726a8080bee253d88921d90252204aa3e4c",
2065
+ "hendrycks_sociology_PL_test.jsonl": "dd56860da95ed793943c9c201f9aa8f5aa8d83753b1bbcbc440fff7ac30f881a",
2066
+ "hendrycks_us_foreign_policy_PL_test.jsonl": "95183b3fcd02d17f5bb7d359b709b3fde13bf6f0f0f1375ab676b0e7dda36786",
2067
+ "hendrycks_virology_PL_test.jsonl": "99e2c6f80907315317619f8c92f8d501ae484619c1169e36d9ad8fa59abfbda7",
2068
+ "hendrycks_world_religions_PL_test.jsonl": "4631dbc58d379cbc46751515f36af682f8bf53db36f3345dbe090d5a00b7b24a"
2069
+ },
2070
+ "acc": 0.247116,
2071
+ "acc_ci95": [
2072
+ 0.239852,
2073
+ 0.254237
2074
+ ],
2075
+ "acc_norm": 0.247116,
2076
+ "acc_norm_ci95": [
2077
+ 0.239994,
2078
+ 0.254237
2079
+ ],
2080
+ "split": "test; micro-average over subjects",
2081
+ "correct_primary": 3470
2082
+ }
2083
+ }
2084
  }
2085
  },
2086
  "additional_measurements": {
README.md CHANGED
@@ -31,9 +31,9 @@ The [CPT Track history](https://track.fabryka.ai/run/71cf98ec-7c60-54a8-ac47-0ea
31
 
32
  ## Objective and starting point
33
 
34
- **Funding update — 11 November 2026, 15:29 UTC:** the operator reported **USD 500 received for continued main-base training**. This is an additional funding receipt, not a measurement of the current account balance or a funded final update count. The earlier USD 425 campaign envelope belongs to the historical 12:07 UTC decision. The objective remains maximum useful affordable exposure with constant learning rates after warmup and **no in-run terminal decay**, retaining cost guards and the operational reserve for a recoverable checkpoint, transfer and safe shutdown. Stop at a verified full checkpoint when the real budget or allocation requires it; a 5,000-update boundary is not required. Positions per dollar measure operating efficiency, not model quality.
35
 
36
- **Operator-recipe snapshot — 11 November 2026, 15:41 UTC:** the approved recipe selects `KROKI=120000`, moving configured WSD decay to update **96,000**. The paired recipe files had been installed on the workstation; the running trainer still used 80,000 until its next restart. File installation is not evidence that a running process has adopted the new schedule. Separately, Track's displayed **80,000-update plan** is a revisable budget estimate, not the trainer's scheduling argument or a promised endpoint. Neither 80,000 nor 120,000 updates, a completion date or over-100B globally unique exposure is guaranteed.
37
 
38
  A separate, continuously curated option targets approximately **10B tokens of quality-selected “gold” data at context 4,096**. It remains isolated from main-base inputs and evaluation material, with its own manifest and licensing/admission gate; books selected with Arkadiusz Słota require title-level license approval. A later finishing run with decay needs its own run and budget decision. Preliminary exclusion lists are not a complete packed 10B corpus or a release, and quality-based selection is not evidence of a measured quality gain. See the [data card](DATA_CARD.md) for the dated inventory and isolation boundary.
39
 
@@ -43,16 +43,16 @@ The unchanged architecture is specified as **1,073,912,851 parameters**, 20 laye
43
 
44
  Architecture and tokenizer lineage includes [SlayerLab/Slayer1B-1.07](https://huggingface.co/SlayerLab/Slayer1B-1.07/blob/a090b54e658fb57dc5166c1fcc90cc6bf5c80106/README.md). The selected 1.08 base and the separate historical 1.08b ablation are different checkpoints; the continuation described here starts from 1.08.
45
 
46
- The [training configuration](TRAINING_CONFIGURATION.md) distinguishes the approved 120,000-update scheduling recipe, its pending runtime adoption at the dated snapshot, Track's separate budget estimate and the actual budget-limited final length. Completed 1.08 and the abandoned first attempt remain separate historical evidence.
47
 
48
  ## Current CPT configuration and budget-limited endpoint
49
 
50
  | Setting | Operator-selected recipe |
51
  |---|---|
52
- | Selected scheduling argument | `KROKI=120000` in the operator-approved recipe; decay begins at update 96,000 after adoption. At the dated 15:41 UTC snapshot the running process still used 80,000; restart readback is required |
53
  | New-source allocation | **20% PL / 80% EN target**; not a measured cumulative model-language ratio |
54
- | Optimizer / schedule | Muon peak 0.02 plus AdamW peak 0.0006; 2,000-update warmup, then constant learning rates under the main-base plan; the configured final-20% WSD decay to 0.00006 is not to be entered in this run |
55
- | Checkpoints | Full state every 200 updates at the verified 11 November, 15:25 UTC recovery; weights-only checkpoints at each 5,000-update boundary; full-state and weights-only workstation copies verified through update 25,000 |
56
  | Code | Launcher v5 SHA256 `796c71016c1e167bfd3ca98b78518f91c3502a96eac233dd7186ef6639fcc656`; exact command and converter pins in the configuration document |
57
 
58
  Operators must review actual available funds, prices and allocation changes without assuming that the decay boundary is permanently unaffordable. An extended run must be reviewed before entering decay; do not shorten `KROKI` to a budget estimate and bring decay forward. Runtime changes need the operator's approved restart procedure and receipt. The trainer and v5 launcher are unchanged; the operator wrapper's argument ceiling was separately approved for the new recipe.
@@ -97,7 +97,7 @@ The last received public sample before the pause is update **450**, training los
97
 
98
  Fresh numerical control and step-zero `INIT-FROM` preflight were verified for the CPT start described above. The earlier permission for the abandoned slot was not reused as that evidence. Any later physical-host restart still requires its own numerical control; successful start does not prove a checkpoint restore.
99
 
100
- Working checkpoints are copied to a trusted workstation, not published as Hugging Face weights. Operators independently verified weights-only copies at updates 5,000, 10,000, 15,000, 20,000 and 25,000, and a separate full-state workstation copy at update 25,000. These transfer records are distinct from the newer recovery checkpoints retained on the running node and from the evaluated updates in the benchmark table. Weights-only copies are retained, while the full workstation recovery copy is replaced by the next verified transfer. This retention policy does not eliminate exposure to a cloud-disk failure between transfers.
101
 
102
  ## Data and exposure
103
 
@@ -107,29 +107,29 @@ The PL/EN balance will follow the approved fresh-data inventory. The user explic
107
 
108
  New material requires licensing, PII filtering, contamination scans, grouped held-outs and deduplication against the base model's full data ancestry. A 1% document holdout is not necessarily 1% of tokens. Historical exposure in the starting weights cannot be undone by filtering new inputs; retrospective benchmark hits must be handled explicitly in evaluation denominators.
109
 
110
- ## Standard benchmark measurements — updates 5,000–20,000
111
 
112
- We thank **BenchTeam / Fabryka AI** for benchmark implementations, protocol work and evaluation proposals. Its independent [full benchmark report](https://github.com/slayerlabs/benchmarks-fabryka-ai/blob/8b547007526971bd2fc35dfe56ff2c7d9b802568/research/results/full-run-2026-11/README.md) is distinct from the **SlayerLab measurements below**, published in the [CPT Track history](https://track.fabryka.ai/run/71cf98ec-7c60-54a8-ac47-0ea7a23f573a). Updates 5,000, 10,000, 15,000 and 20,000 are distinct intermediate working checkpoints of the ongoing 1.09 run, not its eventual budget-limited endpoint or inference-ready weight releases.
113
 
114
- The same nine full splits were scored zero-shot in float32 for 1.07 update 22,327, 1.08 update 1,709 and 1.09 updates 5,000, 10,000, 15,000 and 20,000. Evaluator SHA, dataset revisions, file digests, task protocols and sample counts match across these rows. Scores are percentages with **pointwise 95% confidence intervals**; the named metric is accuracy (`acc`) or length-normalized accuracy (`acc_norm`). WinoGrande uses partial-sentence likelihood; MMLU-PL is micro-averaged. Polish HellaSwag and MMLU use the OpenGPTX task definitions.
115
 
116
- | Task | Metric | n | 1.07 | 1.08 | 1.09, update 5,000 | 1.09, update 10,000 | 1.09, update 15,000 | 1.09, update 20,000 |
117
- |---|---|---:|---:|---:|---:|---:|---:|---:|
118
- | ARC-Challenge | acc_norm | 1,172 | 27.30 [24.74, 29.86] | 30.03 [27.39, 32.68] | 29.35 [26.79, 32.00] | 30.46 [27.82, 33.11] | 32.51 [29.86, 35.24] | 33.11 [30.46, 35.84] |
119
- | HellaSwag | acc_norm | 10,042 | 43.38 [42.43, 44.34] | 43.42 [42.43, 44.39] | 46.53 [45.55, 47.52] | 48.93 [47.97, 49.91] | 49.77 [48.80, 50.76] | 50.38 [49.39, 51.35] |
120
- | PIQA | acc_norm | 1,838 | 66.97 [64.80, 69.15] | 66.21 [64.15, 68.34] | 68.28 [66.21, 70.40] | 68.99 [66.92, 71.11] | 68.99 [66.87, 71.06] | 70.78 [68.72, 72.85] |
121
- | WinoGrande | acc | 1,267 | 53.59 [50.83, 56.35] | 53.67 [50.91, 56.43] | 54.93 [52.17, 57.70] | 54.93 [52.17, 57.62] | 56.75 [54.06, 59.43] | 55.01 [52.25, 57.77] |
122
- | BoolQ | acc | 3,270 | 58.29 [56.61, 59.97] | 59.97 [58.26, 61.59] | 61.80 [60.18, 63.49] | 61.13 [59.45, 62.78] | 59.88 [58.20, 61.53] | 51.41 [49.69, 53.12] |
123
- | OpenBookQA | acc_norm | 500 | 33.60 [29.40, 37.80] | 34.80 [30.60, 39.00] | 35.60 [31.40, 40.00] | 34.40 [30.20, 38.60] | 36.00 [31.80, 40.20] | 35.20 [31.00, 39.40] |
124
- | Belebele, Polish | acc | 900 | 26.00 [23.11, 28.89] | 24.00 [21.22, 26.89] | 24.11 [21.33, 26.89] | 24.11 [21.33, 27.00] | 22.56 [19.89, 25.33] | 24.67 [21.89, 27.56] |
125
- | HellaSwag-PL | acc_norm | 10,042 | 43.83 [42.84, 44.79] | 43.90 [42.90, 44.87] | 44.73 [43.76, 45.71] | 45.51 [44.51, 46.48] | 44.18 [43.21, 45.16] | 44.24 [43.26, 45.22] |
126
- | MMLU-PL | acc | 14,042 | 24.01 [23.30, 24.71] | 23.04 [22.34, 23.74] | 24.10 [23.39, 24.82] | 23.32 [22.62, 24.03] | 23.50 [22.80, 24.21] | 23.86 [23.17, 24.58] |
127
 
128
  Intervals use the percentile bootstrap with 10,000 replicates and seed 20261110. The updated approved aggregate does not supply the random-stream derivation or CI-generator SHA; the previous aggregate documented model-specific streams. **Items are treated as independent; subject, template and source clusters are not modeled.** These are intervals for individual results, not paired differences or training-seed variance; overlap is not a test of a difference.
129
 
130
  The [machine-readable aggregate record](EVALUATION.json) supplies exact values, intervals, counts, dataset revisions and file digests, checkpoint and evaluator identities, and metric selection, without prompts, answers or per-item predictions. ARC-Challenge, OpenBookQA and MMLU-PL use test splits; HellaSwag, PIQA, WinoGrande, BoolQ and HellaSwag-PL use validation splits; Belebele uses the full Polish evaluation set.
131
 
132
- The approved aggregate SHA-256 is `fc6f67a20bdc6ada687349d35de26f6811e27766dfe1aa197082000c03535796`. Its earlier 1.07, 1.08 and 1.09 update-5,000 scores, intervals and dataset pins match the preceding aggregate; the three later checkpoints extend that comparison without replacing the independent two-task protocol below.
133
 
134
  **Descriptive results; contamination scan of the training lineage against these tasks is pending.** Belebele and MMLU-PL are near their four-choice 25% random baseline. No overall average, ranking or release threshold is inferred. Numbers from different evaluators must not be silently substituted: prompting, BOS handling, normalization and software versions can change the result.
135
 
@@ -137,7 +137,7 @@ There is no replay-only control for this continuation, and initialization, expos
137
 
138
  ## Additional GSM8K and SST-2 measurements — update 10,000
139
 
140
- SlayerLab measured this two-task subset on 11 November 2026 using the [pinned BenchTeam code](https://github.com/slayerlabs/benchmarks-fabryka-ai/tree/8b547007526971bd2fc35dfe56ff2c7d9b802568), lm-eval 0.4.9.1, the `slayer` adapter, float32, batch 1 and context limit 4,096. These are full evaluated splits, not a full 50-task run or a copy of the community's independent report. **This separate two-task protocol measures 1.09 update 10,000 only; it does not supply GSM8K or SST-2 results for updates 15,000 or 20,000.** Update 10,000 is an internal working checkpoint, not a published or inference-ready 1.09 weight release.
141
 
142
  | Task | Metric / prompting | n | 1.07 | 1.08 | 1.09, update 10,000 |
143
  |---|---|---:|---:|---:|---:|
 
31
 
32
  ## Objective and starting point
33
 
34
+ **Funding update — 11 November 2026, 15:29 UTC:** the operator reported **USD 500 received for continued main-base training**. This is an additional funding receipt, not a measurement of the current account balance or a funded final update count. The earlier USD 425 campaign envelope belongs to the historical 12:07 UTC decision. The objective remains maximum useful affordable exposure with constant learning rates after warmup and no planned in-run terminal decay, retaining cost guards and the operational reserve for a recoverable checkpoint, transfer and safe shutdown. An extension approaching the configured decay boundary requires review. Stop at a verified full checkpoint when the real budget or allocation requires it; a 5,000-update boundary is not required. Positions per dollar measure operating efficiency, not model quality.
35
 
36
+ **Operator-recipe history and confirmed adoption — 11 November 2026:** at **15:41 UTC**, the paired recipe files selecting `KROKI=120000` had been installed on the workstation while the running trainer still used 80,000. Operators subsequently confirmed adoption at **16:23:47 UTC**, resuming from **update 29,800 on H200** with actual runtime arguments **`--steps 120000 --ckpt-every 200`**, constant LR **0.0006** and configured WSD decay beginning at update **96,000**. This is a dated runtime-adoption receipt, not a claim of current hardware availability. Separately, Track's displayed **80,000-update plan** is a revisable budget estimate, not the trainer's scheduling argument or a promised funded endpoint. Neither 80,000 nor 120,000 updates, a completion date or over-100B globally unique exposure is guaranteed.
37
 
38
  A separate, continuously curated option targets approximately **10B tokens of quality-selected “gold” data at context 4,096**. It remains isolated from main-base inputs and evaluation material, with its own manifest and licensing/admission gate; books selected with Arkadiusz Słota require title-level license approval. A later finishing run with decay needs its own run and budget decision. Preliminary exclusion lists are not a complete packed 10B corpus or a release, and quality-based selection is not evidence of a measured quality gain. See the [data card](DATA_CARD.md) for the dated inventory and isolation boundary.
39
 
 
43
 
44
  Architecture and tokenizer lineage includes [SlayerLab/Slayer1B-1.07](https://huggingface.co/SlayerLab/Slayer1B-1.07/blob/a090b54e658fb57dc5166c1fcc90cc6bf5c80106/README.md). The selected 1.08 base and the separate historical 1.08b ablation are different checkpoints; the continuation described here starts from 1.08.
45
 
46
+ The [training configuration](TRAINING_CONFIGURATION.md) distinguishes the approved 120,000-update scheduling recipe, its confirmed runtime adoption, Track's separate budget estimate and the actual budget-limited final length. Completed 1.08 and the abandoned first attempt remain separate historical evidence.
47
 
48
  ## Current CPT configuration and budget-limited endpoint
49
 
50
  | Setting | Operator-selected recipe |
51
  |---|---|
52
+ | Selected scheduling argument | `KROKI=120000`; confirmed runtime `--steps 120000` adoption at 11 November, 16:23:47 UTC, resuming from update 29,800 on H200; configured decay begins at update 96,000 |
53
  | New-source allocation | **20% PL / 80% EN target**; not a measured cumulative model-language ratio |
54
+ | Optimizer / schedule | Muon peak 0.02 plus AdamW peak 0.0006; 2,000-update warmup, then constant learning rates under the main-base plan; adoption receipt confirms LR 0.0006. The configured final-20% WSD decay to 0.00006 remains at update 96,000; an extension approaching it requires review |
55
+ | Checkpoints | Runtime `--ckpt-every 200` confirmed at the 16:23:47 UTC adoption; weights-only checkpoints at each 5,000-update boundary; full-state and weights-only workstation copies verified through update 25,000 |
56
  | Code | Launcher v5 SHA256 `796c71016c1e167bfd3ca98b78518f91c3502a96eac233dd7186ef6639fcc656`; exact command and converter pins in the configuration document |
57
 
58
  Operators must review actual available funds, prices and allocation changes without assuming that the decay boundary is permanently unaffordable. An extended run must be reviewed before entering decay; do not shorten `KROKI` to a budget estimate and bring decay forward. Runtime changes need the operator's approved restart procedure and receipt. The trainer and v5 launcher are unchanged; the operator wrapper's argument ceiling was separately approved for the new recipe.
 
97
 
98
  Fresh numerical control and step-zero `INIT-FROM` preflight were verified for the CPT start described above. The earlier permission for the abandoned slot was not reused as that evidence. Any later physical-host restart still requires its own numerical control; successful start does not prove a checkpoint restore.
99
 
100
+ Working checkpoints are not published as Hugging Face weights. Operators independently verified weights-only workstation copies at updates 5,000, 10,000, 15,000, 20,000 and 25,000, and a separate full-state workstation copy at update 25,000. The later full 30,000 transfer was interrupted in `.tmp`; benchmark conversion is not evidence of a completed full-state workstation copy. These transfer records are distinct from newer recovery checkpoints on the node and from evaluated working updates through 30,000 below. Weights-only copies are retained, while the full workstation recovery copy is replaced by the next verified transfer. This retention policy does not eliminate exposure to a cloud-disk failure between transfers.
101
 
102
  ## Data and exposure
103
 
 
107
 
108
  New material requires licensing, PII filtering, contamination scans, grouped held-outs and deduplication against the base model's full data ancestry. A 1% document holdout is not necessarily 1% of tokens. Historical exposure in the starting weights cannot be undone by filtering new inputs; retrospective benchmark hits must be handled explicitly in evaluation denominators.
109
 
110
+ ## Standard benchmark measurements — updates 5,000–30,000
111
 
112
+ We thank **BenchTeam / Fabryka AI** for benchmark implementations, protocol work and evaluation proposals. Its independent [full benchmark report](https://github.com/slayerlabs/benchmarks-fabryka-ai/blob/8b547007526971bd2fc35dfe56ff2c7d9b802568/research/results/full-run-2026-11/README.md) is distinct from the **SlayerLab measurements below**, published in the [CPT Track history](https://track.fabryka.ai/run/71cf98ec-7c60-54a8-ac47-0ea7a23f573a). Updates 5,000, 10,000, 15,000, 20,000, 25,000 and 30,000 are distinct intermediate working checkpoints of the ongoing 1.09 run, not its eventual budget-limited endpoint or inference-ready weight releases.
113
 
114
+ The same nine full splits were scored zero-shot in float32 for 1.07 update 22,327, 1.08 update 1,709 and 1.09 updates 5,000, 10,000, 15,000, 20,000, 25,000 and 30,000. Evaluator SHA, dataset revisions, file digests, task protocols and sample counts match across these rows. Scores are percentages with **pointwise 95% confidence intervals**; the named metric is accuracy (`acc`) or length-normalized accuracy (`acc_norm`). WinoGrande uses partial-sentence likelihood; MMLU-PL is micro-averaged. Polish HellaSwag and MMLU use the OpenGPTX task definitions.
115
 
116
+ | Task | Metric | n | 1.07 | 1.08 | 1.09, update 5,000 | 1.09, update 10,000 | 1.09, update 15,000 | 1.09, update 20,000 | 1.09, update 25,000 | 1.09, update 30,000 |
117
+ |---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|
118
+ | ARC-Challenge | acc_norm | 1,172 | 27.30 [24.74, 29.86] | 30.03 [27.39, 32.68] | 29.35 [26.79, 32.00] | 30.46 [27.82, 33.11] | 32.51 [29.86, 35.24] | 33.11 [30.46, 35.84] | 33.79 [31.06, 36.52] | 33.02 [30.29, 35.67] |
119
+ | HellaSwag | acc_norm | 10,042 | 43.38 [42.43, 44.34] | 43.42 [42.43, 44.39] | 46.53 [45.55, 47.52] | 48.93 [47.97, 49.91] | 49.77 [48.80, 50.76] | 50.38 [49.39, 51.35] | 51.17 [50.18, 52.14] | 52.00 [51.02, 52.98] |
120
+ | PIQA | acc_norm | 1,838 | 66.97 [64.80, 69.15] | 66.21 [64.15, 68.34] | 68.28 [66.21, 70.40] | 68.99 [66.92, 71.11] | 68.99 [66.87, 71.06] | 70.78 [68.72, 72.85] | 71.55 [69.48, 73.61] | 70.67 [68.61, 72.74] |
121
+ | WinoGrande | acc | 1,267 | 53.59 [50.83, 56.35] | 53.67 [50.91, 56.43] | 54.93 [52.17, 57.70] | 54.93 [52.17, 57.62] | 56.75 [54.06, 59.43] | 55.01 [52.25, 57.77] | 55.88 [53.12, 58.64] | 57.14 [54.46, 59.83] |
122
+ | BoolQ | acc | 3,270 | 58.29 [56.61, 59.97] | 59.97 [58.26, 61.59] | 61.80 [60.18, 63.49] | 61.13 [59.45, 62.78] | 59.88 [58.20, 61.53] | 51.41 [49.69, 53.12] | 52.20 [50.52, 53.91] | 60.67 [58.99, 62.35] |
123
+ | OpenBookQA | acc_norm | 500 | 33.60 [29.40, 37.80] | 34.80 [30.60, 39.00] | 35.60 [31.40, 40.00] | 34.40 [30.20, 38.60] | 36.00 [31.80, 40.20] | 35.20 [31.00, 39.40] | 37.20 [32.80, 41.40] | 36.20 [32.00, 40.40] |
124
+ | Belebele, Polish | acc | 900 | 26.00 [23.11, 28.89] | 24.00 [21.22, 26.89] | 24.11 [21.33, 26.89] | 24.11 [21.33, 27.00] | 22.56 [19.89, 25.33] | 24.67 [21.89, 27.56] | 25.33 [22.44, 28.22] | 24.44 [21.67, 27.33] |
125
+ | HellaSwag-PL | acc_norm | 10,042 | 43.83 [42.84, 44.79] | 43.90 [42.90, 44.87] | 44.73 [43.76, 45.71] | 45.51 [44.51, 46.48] | 44.18 [43.21, 45.16] | 44.24 [43.26, 45.22] | 43.94 [42.94, 44.91] | 45.75 [44.77, 46.73] |
126
+ | MMLU-PL | acc | 14,042 | 24.01 [23.30, 24.71] | 23.04 [22.34, 23.74] | 24.10 [23.39, 24.82] | 23.32 [22.62, 24.03] | 23.50 [22.80, 24.21] | 23.86 [23.17, 24.58] | 24.26 [23.57, 24.96] | 24.71 [23.99, 25.42] |
127
 
128
  Intervals use the percentile bootstrap with 10,000 replicates and seed 20261110. The updated approved aggregate does not supply the random-stream derivation or CI-generator SHA; the previous aggregate documented model-specific streams. **Items are treated as independent; subject, template and source clusters are not modeled.** These are intervals for individual results, not paired differences or training-seed variance; overlap is not a test of a difference.
129
 
130
  The [machine-readable aggregate record](EVALUATION.json) supplies exact values, intervals, counts, dataset revisions and file digests, checkpoint and evaluator identities, and metric selection, without prompts, answers or per-item predictions. ARC-Challenge, OpenBookQA and MMLU-PL use test splits; HellaSwag, PIQA, WinoGrande, BoolQ and HellaSwag-PL use validation splits; Belebele uses the full Polish evaluation set.
131
 
132
+ The approved aggregate SHA-256 is `2d257a95523605b06aa964cb974f65d86ab025875cc1f87dc98b242b90729440`. All six earlier model records retain their scores, intervals, counts and dataset pins; the distinct update-25,000 and update-30,000 records extend that comparison without replacing the independent two-task protocol below.
133
 
134
  **Descriptive results; contamination scan of the training lineage against these tasks is pending.** Belebele and MMLU-PL are near their four-choice 25% random baseline. No overall average, ranking or release threshold is inferred. Numbers from different evaluators must not be silently substituted: prompting, BOS handling, normalization and software versions can change the result.
135
 
 
137
 
138
  ## Additional GSM8K and SST-2 measurements — update 10,000
139
 
140
+ SlayerLab measured this two-task subset on 11 November 2026 using the [pinned BenchTeam code](https://github.com/slayerlabs/benchmarks-fabryka-ai/tree/8b547007526971bd2fc35dfe56ff2c7d9b802568), lm-eval 0.4.9.1, the `slayer` adapter, float32, batch 1 and context limit 4,096. These are full evaluated splits, not a full 50-task run or a copy of the community's independent report. **This separate two-task protocol measures 1.09 update 10,000 only; it does not supply GSM8K or SST-2 results for updates 15,000, 20,000, 25,000 or 30,000.** Update 10,000 is an internal working checkpoint, not a published or inference-ready 1.09 weight release.
141
 
142
  | Task | Metric / prompting | n | 1.07 | 1.08 | 1.09, update 10,000 |
143
  |---|---|---:|---:|---:|---:|
TRAINING_CONFIGURATION.md CHANGED
@@ -13,7 +13,7 @@ The selected base is [SlayerLab/Slayer1B-1.08 at 5f854a0cc7160594d54d1788d37a44e
13
 
14
  The approved standalone c01 reference has TRAIN SHA256 `3341080241aeabd1e03c77530a6d735cb9501ecd788fba60ea065a944cd09ee1` and occupies input range `[0, 5,347,737,900)`. Its initial one-slot queue snapshot has SHA256 `48144a407bcd3b2da8ad9106e0f8d94eba4335bf08dcbb0492f99acbdce91d98`. Separately approved appended queues may extend it while preserving the existing prefix; these initial digests identify the original snapshots, not the whole growing file after an append. **`D/val.bin` is the fresh CL validation file**, SHA256 `4c50400b983b41411ebf2f68c09b1ab554517a99af270b731a84c80f4a2636a1`, not the legacy Logika validation in the construction output. Operators verified these identities for the initial launch; each later append still needs its own approved byte checks.
15
 
16
- The archived v3 invocation later in this document is historical. The v5 example below records the selected 120,000-update mechanical scheduling recipe, not a financial commitment to run that many updates. At the 11 November, 15:41 UTC operator snapshot, the paired recipe files were installed on the workstation but the running trainer still used 80,000 until restart. Installation, runtime adoption and completed exposure are separate receipts. Cost monitoring, availability checks and verified checkpoint transfers remain required.
17
 
18
  ## Operator-approved v5 recipe and endpoint selection
19
 
@@ -34,11 +34,11 @@ PY=/workspace/venv/bin/python TORCHRUN=/workspace/venv/bin/torchrun \
34
  bash /workspace/kod/uruchom_109.sh
35
  ```
36
 
37
- Set `START`, `GPU`, `S`, `KROKI` and `CKPT_EVERY` explicitly. The operator-approved recipe selects **`KROKI=120000`**, moving the 20% WSD decay boundary to **update 96,000**. W4 approved the paired argument-file and operator-wrapper ceiling change on 11 November at 15:40 UTC; the operator installed them on the workstation at 15:41 UTC. The trainer and v5 launcher code are unchanged. The still-running 80,000-update process at that snapshot retained its earlier update-64,000 decay boundary until restart; read back the changed argument after the approved restart before treating the new schedule as active.
38
 
39
  The launcher permits this change only before `min(0.8 × old steps, 0.8 × new steps)`: for 80,000 → 120,000, before **update 64,000**. Review actual available funds, prices and allocation availability without assuming any decay boundary is permanently unaffordable. If more funding or lower prices make update 96,000 reachable, review the plan before it is reached to preserve the main-base no-decay objective. Do not shorten `KROKI` to a budget estimate and move decay earlier. Track's displayed **80,000-update plan** is a separate revisable estimate, not this scheduling argument or a guaranteed endpoint.
40
 
41
- Control values remain microbatch 8 × accumulation 4 × eight ranks, context 4,096, Muon peak 0.02 / AdamW peak 0.0006 and 2,000-update warmup. The configured final 20% decay to 0.00006 remains in the scheduler but is not the main-base objective. Full-state checkpoints used 400-update intervals historically; the verified 11 November, 15:25 UTC recovery used **200**, following the earlier cadence decision for short spot allocations. Each 5,000-update boundary also retains a weights-only checkpoint. Cadence is an operational choice recorded separately from the learning-rate schedule. Target source allocation is 20% PL / 80% EN, not the cumulative model-language ratio.
42
 
43
  `--init-from` imports weights/config from 1.08 into this new experiment with a fresh optimizer. `--resume` restores this experiment's own state; it must not resume the abandoned random-initialized attempt. A verified new run and a verified recovery are separate receipts.
44
 
 
13
 
14
  The approved standalone c01 reference has TRAIN SHA256 `3341080241aeabd1e03c77530a6d735cb9501ecd788fba60ea065a944cd09ee1` and occupies input range `[0, 5,347,737,900)`. Its initial one-slot queue snapshot has SHA256 `48144a407bcd3b2da8ad9106e0f8d94eba4335bf08dcbb0492f99acbdce91d98`. Separately approved appended queues may extend it while preserving the existing prefix; these initial digests identify the original snapshots, not the whole growing file after an append. **`D/val.bin` is the fresh CL validation file**, SHA256 `4c50400b983b41411ebf2f68c09b1ab554517a99af270b731a84c80f4a2636a1`, not the legacy Logika validation in the construction output. Operators verified these identities for the initial launch; each later append still needs its own approved byte checks.
15
 
16
+ The archived v3 invocation later in this document is historical. The v5 example below records the selected 120,000-update mechanical scheduling recipe, not a financial commitment to run that many updates. The paired files were installed on the workstation on 11 November at 15:41 UTC; independent readbacks confirmed adoption after `RESUME @ 29800` at **16:23:47 UTC**. Installation, runtime adoption and completed exposure remain separate receipts. That recovery is not a claim of uninterrupted hardware availability. Cost monitoring, availability checks and verified checkpoint transfers remain required.
17
 
18
  ## Operator-approved v5 recipe and endpoint selection
19
 
 
34
  bash /workspace/kod/uruchom_109.sh
35
  ```
36
 
37
+ Set `START`, `GPU`, `S`, `KROKI` and `CKPT_EVERY` explicitly. The operator-approved recipe selects **`KROKI=120000`**, moving the 20% WSD decay boundary to **update 96,000**. The security reviewer approved the paired argument-file and operator-wrapper ceiling change on 11 November at 15:40 UTC; the operator installed them on the workstation at 15:41 UTC. Adoption was then independently verified after the **16:23:47 UTC restart from update 29,800**: runtime `--steps 120000`, `--ckpt-every 200` and learning rate **0.0006**, with the change from 80,000 to 120,000 recorded by the launcher. The trainer and v5 launcher code are unchanged. This confirms the approved schedule in that recovery, not current hardware availability or a funded final update count.
38
 
39
  The launcher permits this change only before `min(0.8 × old steps, 0.8 × new steps)`: for 80,000 → 120,000, before **update 64,000**. Review actual available funds, prices and allocation availability without assuming any decay boundary is permanently unaffordable. If more funding or lower prices make update 96,000 reachable, review the plan before it is reached to preserve the main-base no-decay objective. Do not shorten `KROKI` to a budget estimate and move decay earlier. Track's displayed **80,000-update plan** is a separate revisable estimate, not this scheduling argument or a guaranteed endpoint.
40
 
41
+ Control values remain microbatch 8 × accumulation 4 × eight ranks, context 4,096, Muon peak 0.02 / AdamW peak 0.0006 and 2,000-update warmup. The configured final 20% decay to 0.00006 remains in the scheduler but is not the main-base objective. Full-state checkpoints used 400-update intervals historically; the verified 11 November, 14:53 UTC recovery used **200**, following the earlier cadence decision for short spot allocations. Each 5,000-update boundary also retains a weights-only checkpoint. Cadence is an operational choice recorded separately from the learning-rate schedule. Target source allocation is 20% PL / 80% EN, not the cumulative model-language ratio.
42
 
43
  `--init-from` imports weights/config from 1.08 into this new experiment with a fresh optimizer. `--resume` restores this experiment's own state; it must not resume the abandoned random-initialized attempt. A verified new run and a verified recovery are separate receipts.
44