examples.jsonl

{"id": "gollem-v5-64m-byteppl", "repo_id": "SlayerLab/gollem-v5-ckpts", "metric": "wikitext2_byte_ppl", "board_value": 2.016, "board_bpb": 1.012, "rebench_value": 2.370, "rebench_bpb": 1.245, "correct_value": 2.372, "correct_bpb": 1.246, "bpt_factor_wrong": 4.755, "bpt_factor_correct": 3.8605, "raw_text_bytes": 1292008, "token_count": 334674, "tokenizer": "BPE-12288", "note": "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.", "verified": true}
{"id": "gollem-v5-64m-arc-control", "repo_id": "SlayerLab/gollem-v5-ckpts", "metric": "arc_easy", "board_value": 47.94, "rebench_value": 47.94, "delta": 0.0, "protocol": "bare-prompt, 256-token clip (Glint board)", "note": "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.", "verified": true}
{"id": "gollem-v5-64m-blimp", "repo_id": "SlayerLab/gollem-v5-ckpts", "metric": "blimp", "board_value": 77.84, "rebench_value": 75.99, "delta": -1.85, "note": "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).", "verified": true}
{"id": "anka-50m-tied-dedup", "repo_id": "Melih1234/ANKA-50M-RMW3", "metric": "param_count", "card_claim": "48,944,657 unique", "header_total": 57333265, "tied_copy_subtracted": 8388608, "vocab_size": 16384, "hidden_size": 512, "result": "clean", "note": "Card unique count = header total minus one tied-embedding copy (16384x512=8,388,608). Exact. 199 BF16 tensors.", "verified": true}
{"id": "nexi-g1-ppl", "repo_id": "aksern/nexi-g1", "metric": "perplexity", "param_count": 30339456, "vocab_size": 50257, "ppl_tinystories_valid_2000": 43.56, "ppl_wikitext2_test": 157.53, "note": "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).", "verified": true}
{"id": "char-gpt-1.2m-heldout", "repo_id": "Compactbot/char-gpt-1.2m", "metric": "heldout_perplexity", "param_count": 1216000, "vocab_size": 65, "test_loss": 1.4369, "test_perplexity": 4.21, "test_tokens": 49674, "note": "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.", "verified": true}