{ "step": 76294, "parameters": 148910738, "dataset_revisions": { "Salesforce/wikitext": "b08601e04326c79dfdd32d625aee71d232d685c3", "nyu-mll/blimp": "877fba0801ffb7cbd8c39c1ff314a46f053f6036", "allenai/ai2_arc": "210d026faf9955653af8916fad021475a3f00453" }, "continuation_tokens": 20000014336, "provenance": { "source_repo": "SlayerLab/gollem-v5-ckpts", "source_revision": "963440ce6ab4ada7e95da4c1faa28ebb00c082d4", "source_file": "final_128m_16x768/ckpt_760k.pt", "source_sha256": "95d8b43fcd9c5fa06566da13dd03856b00300bad6f37fb4dc9442c18be9ac372", "source_steps": 760000, "source_training_tokens": 24903680000, "parameters": 148910738, "checks": [ { "length": 16, "max_absolute_logit_error": 0.0 }, { "length": 128, "max_absolute_logit_error": 0.0 }, { "length": 256, "max_absolute_logit_error": 0.0 } ], "optimizer": "reset for continued pretraining", "method": "FFN expansion with zero new down columns; append three identity residual blocks" }, "protocol": "GoLLeM r6 GLINT parity; bf16 forward, fp32 logprob; context 256", "wikitext2_token_ppl": 21.452430290724198, "bytes_per_token": 3.8604970807412586, "wikitext2_byte_ppl": 2.2125733814416444, "blimp_acc": 79.47, "blimp_n": 67000, "arc_easy_acc": 54.97, "arc_n": 2376, "efficiency_fixed_board_snapshot": 77.1358484476474, "leader_reported_efficiency_snapshot": 80.20023332743875, "comparison_caveat": "Competitor scores are self-reported, not independently re-evaluated under this protocol" }