id string | repo_id string | metric string | board_value float64 | board_bpb float64 | rebench_value float64 | rebench_bpb float64 | correct_value float64 | correct_bpb float64 | bpt_factor_wrong float64 | bpt_factor_correct float64 | raw_text_bytes int64 | token_count int64 | tokenizer string | note string | verified bool | delta float64 | protocol string | card_claim string | header_total int64 | tied_copy_subtracted int64 | vocab_size int64 | hidden_size int64 | result string | param_count int64 | ppl_tinystories_valid_2000 float64 | ppl_wikitext2_test float64 | test_loss float64 | test_perplexity float64 | test_tokens int64 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
gollem-v5-64m-byteppl | SlayerLab/gollem-v5-ckpts | wikitext2_byte_ppl | 2.016 | 1.012 | 2.37 | 1.245 | 2.372 | 1.246 | 4.755 | 3.8605 | 1,292,008 | 334,674 | BPE-12288 | Card's 2.016/1.012 used a wrong bytes-per-token factor (4.755). Correct BPT = raw_bytes/tokens = 3.8605, giving byte_ppl 2.372 / BPB 1.246, matching an independent re-benchmark (decode-each-token-to-UTF8 = 3.864 BPT). Verified 2026-09-25. | true | null | null | null | null | null | null | null | null | null | null | null | null | null | null |
gollem-v5-64m-arc-control | SlayerLab/gollem-v5-ckpts | arc_easy | 47.94 | null | 47.94 | null | null | null | null | null | null | null | null | ARC-Easy matched to 2 decimals = load control proving checkpoint loads + loglikelihood scoring sound. BLiMP gap (77.84 vs 75.99) and WikiText gap are convention differences, not load bugs. | true | 0 | bare-prompt, 256-token clip (Glint board) | null | null | null | null | null | null | null | null | null | null | null | null |
gollem-v5-64m-blimp | SlayerLab/gollem-v5-ckpts | blimp | 77.84 | null | 75.99 | null | null | null | null | null | null | null | null | Re-benchmark used bare-prompt 256-clip protocol. Gap attributed to harness scoring difference; board defers to its documented protocol. Honest eff ~75.9 (not 77.51). | true | -1.85 | null | null | null | null | null | null | null | null | null | null | null | null | null |
anka-50m-tied-dedup | Melih1234/ANKA-50M-RMW3 | param_count | null | null | null | null | null | null | null | null | null | null | null | Card unique count = header total minus one tied-embedding copy (16384x512=8,388,608). Exact. 199 BF16 tensors. | true | null | null | 48,944,657 unique | 57,333,265 | 8,388,608 | 16,384 | 512 | clean | null | null | null | null | null | null |
nexi-g1-ppl | aksern/nexi-g1 | perplexity | null | null | null | null | null | null | null | null | null | null | null | GPT-2 BPE (vocab 50257). TinyStories ppl (43.6) is in-distribution; WikiText2 ppl (157.5) is a distribution mismatch (training mix = wikipedia/imdb/marketing/4chan), not a bug. TinyStories eval used first 2000 of 21990 valid rows (subset). | true | null | null | null | null | null | 50,257 | null | null | 30,339,456 | 43.56 | 157.53 | null | null | null |
char-gpt-1.2m-heldout | Compactbot/char-gpt-1.2m | heldout_perplexity | null | null | null | null | null | null | null | null | null | null | null | Char-level LM (65-char vocab, TinyStories). Full held-out test split scored with next-token CE. Card's earlier val 1.9046 was a single-epoch 60-batch sample; corrected to full-split 4.21. Char models that cannot read benchmark text should report held-out PPL, not a forced suite score. | true | null | null | null | null | null | 65 | null | null | 1,216,000 | null | null | 1.4369 | 4.21 | 49,674 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.