{ "size": "700m", "version": "v0.1.0", "run": "f10_slp1_uni8k_d700m_plus_all_v2_3x", "torch": "2.14.0+cu130", "transformers": "5.18.0", "hf_load_seconds": 0.4, "hf_param_counts": { "total": 704128512, "embedding": 12288000, "head": 12288000, "non_embedding": 679552512 }, "orig_param_counts": { "total": 704128512, "embedding": 12288000, "token_embedding": 12288000, "position_embedding": 0, "value_embedding": 0, "head": 12288000, "non_embedding": 679552512 }, "tied_head": false, "hf_head_equals_wte": false, "logits": [ { "input": "dcs_gold", "tokens": 19, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.10451602935791016, "mean_abs_diff_vs_fp32_master": 0.017697827890515327, "argmax_agree_vs_fp32_master": 1.0, "logit_abs_max": 13.280336380004883 }, { "input": "gita", "tokens": 27, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.13284730911254883, "mean_abs_diff_vs_fp32_master": 0.0162887554615736, "argmax_agree_vs_fp32_master": 1.0, "logit_abs_max": 24.75530242919922 }, { "input": "ood", "tokens": 512, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.15660560131072998, "mean_abs_diff_vs_fp32_master": 0.018970483914017677, "argmax_agree_vs_fp32_master": 0.99609375, "logit_abs_max": 26.321086883544922 }, { "input": "prose", "tokens": 61, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.13342761993408203, "mean_abs_diff_vs_fp32_master": 0.016097350046038628, "argmax_agree_vs_fp32_master": 1.0, "logit_abs_max": 17.176036834716797 }, { "input": "vedic", "tokens": 22, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.10383224487304688, "mean_abs_diff_vs_fp32_master": 0.016606679186224937, "argmax_agree_vs_fp32_master": 1.0, "logit_abs_max": 17.724536895751953 }, { "input": "random_512", "tokens": 512, "max_abs_diff_vs_same_bf16_weights": 0.0, "mean_abs_diff_vs_same_bf16_weights": 0.0, "max_abs_diff_vs_fp32_master": 0.20844125747680664, "mean_abs_diff_vs_fp32_master": 0.013787541538476944, "argmax_agree_vs_fp32_master": 0.99609375, "logit_abs_max": 20.799428939819336 } ], "logits_max_abs_diff_impl": 0.0, "logits_max_abs_diff_bf16_storage": 0.20844125747680664, "left_padding_max_abs_diff": 3.528594970703125e-05, "tokenizer_ids_identical": "95/100", "tokenizer_first_difference": { "set": "ood", "index": 2 }, "tokenizer_ids_identical_fence_latin_false": "100/100", "tokenizer_roundtrip_exact": "100/100", "bpb_sample": { "records_per_set": 20, "block": 512, "stride": 256, "original_fp32_master": { "pooled": 0.48407, "dcs_gold": 0.66705, "gita": 0.10244, "ood": 0.49614, "prose": 0.48516, "vedic": 0.67513 }, "hf_bf16_weights_fp32_compute": { "pooled": 0.48405, "dcs_gold": 0.66683, "gita": 0.1024, "ood": 0.49612, "prose": 0.48514, "vedic": 0.6754 }, "pooled_rel_diff": -3.880216049873875e-05, "seconds": 104.1 }, "bpb_full": { "note": "every record, HF bf16 weights via transformers + trust_remote_code, extbench_bpb.py --fence_latin_false --no_bos --cfgs 512:256, CUDA bf16; ref = eval/bpb*.json", "std": { "hf_ex_gita": 0.53059, "ref_ex_gita": 0.53067, "rel_diff": -0.00016257954652671732, "hf": { "dcs_gold": 0.55111, "gita": 0.07577, "ood": 0.52853, "prose": 0.51859, "vedic": 0.66687 }, "ref": { "dcs_gold": 0.55126, "gita": 0.07576, "ood": 0.52863, "prose": 0.5186, "vedic": 0.66705 }, "bytes_equal": true }, "clean_v1": { "hf_ex_gita": 0.55404, "ref_ex_gita": 0.55415, "rel_diff": -0.00019493170581936874, "hf": { "dcs_gold": 0.55453, "gita": 0.08184, "ood": 0.55967, "prose": 0.53144, "vedic": 0.66953 }, "ref": { "dcs_gold": 0.55472, "gita": 0.08182, "ood": 0.55979, "prose": 0.53148, "vedic": 0.66967 }, "bytes_equal": true } } }