"""All end-to-end results, transcribed from the full per-task result tables. Downstream rows: (task, raw_mean, s_pretrain, s_eval, s_binom, s_total, target_mean, delta_pp, delta_rel_pct, z). Paloma rows: (corpus, raw_bpb, s_pretrain, s_eval, s_total, target_bpb, delta_bpb, delta_rel_pct, z). """ # -------------------------------------------------------------------------- # Task display names (harness key -> paper label) # -------------------------------------------------------------------------- TASK_LABEL = { "lambada_openai": "LAMBADA", "arc_challenge": "ARC-Challenge", "arc_easy": "ARC-Easy", "hellaswag": "HellaSwag", "mmlu": "MMLU", "copa": "COPA", "boolq": "BoolQ", "sciq": "SciQ", "race": "RACE", "anli_r1": "ANLI-R1", "anli_r2": "ANLI-R2", "anli_r3": "ANLI-R3", "piqa": "PIQA", "truthfulqa_mc2": "TruthfulQA-MC2", "rte": "RTE", "blimp": "BLiMP", "pubmedqa": "PubMedQA", "cb": "CB", "triviaqa": "TriviaQA", "commonsense_qa": "CommonsenseQA", "logiqa": "LogiQA", "openbookqa": "OpenBookQA", "nq_open": "NQ-Open", "winogrande": "WinoGrande", "coqa": "CoQA", "squadv2": "SQuAD v2", "social_iqa": "SocialIQA", } # Grouping used for the heat-map ordering (capability families) TASK_GROUPS = [ ("Language modelling", ["lambada_openai", "blimp"]), ("Science / knowledge", ["sciq", "arc_easy", "arc_challenge", "pubmedqa", "mmlu"]), ("Commonsense", ["hellaswag", "piqa", "commonsense_qa", "social_iqa", "copa", "openbookqa", "winogrande"]), ("Logic / NLI", ["logiqa", "anli_r1", "anli_r2", "anli_r3", "rte", "cb", "boolq", "truthfulqa_mc2"]), ("Reading / open QA", ["race", "coqa", "squadv2", "triviaqa", "nq_open"]), ] PALOMA_LABEL = { "paloma_falcon-refinedweb": "Falcon-RefinedWeb", "paloma_c4_en": "C4-en", "paloma_c4_100_domains": "C4-100-domains", "paloma_dolma_100_subreddits": "Dolma-100-subreddits", "paloma_dolma-v1_5": "Dolma-v1.5", "paloma_m2d2_wikipedia_unsplit": "M2D2-Wikipedia", "paloma_m2d2_s2orc_unsplit": "M2D2-S2ORC", "paloma_mc4": "mC4-en", "paloma_redpajama": "RedPajama", "paloma_wikitext_103": "WikiText-103", "paloma_ptb": "Penn Treebank", } PALOMA_ORDER = [ "paloma_falcon-refinedweb", "paloma_c4_en", "paloma_c4_100_domains", "paloma_mc4", "paloma_dolma_100_subreddits", "paloma_dolma-v1_5", "paloma_redpajama", "paloma_m2d2_wikipedia_unsplit", "paloma_m2d2_s2orc_unsplit", "paloma_wikitext_103", "paloma_ptb", ] PALOMA_FAMILY = { # web / social crawls "paloma_falcon-refinedweb": "web", "paloma_c4_en": "web", "paloma_c4_100_domains": "web", "paloma_mc4": "web", "paloma_dolma_100_subreddits": "web", # mixed-source corpora (web + curated sources) "paloma_dolma-v1_5": "mixed", "paloma_redpajama": "mixed", # curated / edited corpora "paloma_m2d2_wikipedia_unsplit": "curated", "paloma_m2d2_s2orc_unsplit": "curated", "paloma_wikitext_103": "curated", "paloma_ptb": "curated", } # -------------------------------------------------------------------------- # HQ tier # -------------------------------------------------------------------------- HQ_V1 = [ ("lambada_openai", 39.6, 0.311, 0.000, 0.684, 0.311, 41.9, +2.4, +5.98, +7.6), ("arc_challenge", 30.9, 0.215, 0.000, 1.346, 0.215, 30.2, -0.7, -2.16, -3.1), ("hellaswag", 49.3, 0.398, 0.000, 0.499, 0.398, 48.3, -1.0, -1.95, -2.4), ("mmlu", 25.2, 0.716, 0.000, 0.369, 0.716, 26.2, +1.1, +4.16, +1.5), ("copa", 72.3, 1.528, 0.000, 4.424, 1.528, 74.5, +2.2, +3.00, +1.4), ("boolq", 57.4, 1.647, 0.000, 0.861, 1.647, 59.7, +2.3, +4.00, +1.4), ("sciq", 79.0, 0.764, 0.000, 1.273, 0.764, 80.0, +1.0, +1.29, +1.3), ("race", 31.3, 0.829, 0.000, 1.439, 0.829, 31.9, +0.6, +1.99, +0.8), ("anli_r2", 33.8, 0.379, 0.000, 1.496, 0.379, 33.5, -0.3, -0.84, -0.8), ("piqa", 70.2, 0.649, 0.000, 1.070, 0.649, 69.7, -0.5, -0.69, -0.7), ("truthfulqa_mc2", 37.9, 0.704, 0.000, 1.403, 0.704, 38.4, +0.5, +1.34, +0.7), ("rte", 49.5, 1.532, 0.000, 3.009, 1.532, 50.5, +1.1, +2.19, +0.7), ("blimp", 81.3, 0.894, 0.003, 0.133, 0.894, 81.8, +0.6, +0.69, +0.6), ("anli_r3", 34.8, 2.014, 0.000, 1.371, 2.014, 33.8, -1.1, -3.03, -0.5), ("pubmedqa", 54.2, 1.186, 0.000, 2.229, 1.186, 54.7, +0.5, +1.01, +0.5), ("cb", 49.4, 11.619, 0.000, 6.644, 11.619, 54.5, +5.1, +10.24, +0.4), ("triviaqa", 5.6, 1.525, 0.000, 0.175, 1.525, 6.2, +0.6, +11.51, +0.4), ("commonsense_qa", 19.8, 0.855, 0.000, 1.146, 0.855, 20.1, +0.3, +1.65, +0.4), ("arc_easy", 59.6, 0.973, 0.000, 1.005, 0.973, 60.0, +0.4, +0.61, +0.4), ("logiqa", 21.1, 0.900, 0.000, 1.603, 0.900, 21.4, +0.3, +1.58, +0.4), ("openbookqa", 34.6, 0.907, 0.141, 2.132, 0.918, 34.9, +0.3, +0.77, +0.3), ("nq_open", 3.2, 0.215, 0.000, 0.297, 0.215, 3.3, +0.1, +1.70, +0.3), ("winogrande", 53.4, 0.721, 0.000, 1.402, 0.721, 53.3, -0.1, -0.28, -0.2), ("coqa", 23.2, 0.854, 0.000, 1.625, 0.854, 23.1, -0.1, -0.49, -0.1), ("anli_r1", 31.8, 1.480, 0.000, 1.473, 1.480, 31.9, +0.1, +0.47, +0.1), ("squadv2", 256.9, 25.319, 15.350, None, 29.608, 256.2, -0.7, -0.28, -0.0), ("social_iqa", 40.5, 0.795, 0.000, 1.111, 0.795, 40.6, +0.0, +0.02, +0.0), ] HQ_V1_PALOMA = [ ("paloma_ptb", 1.0378, 0.0012, 0.0000, 0.0012, 1.1591, +0.1213, +11.69, +101.9), ("paloma_falcon-refinedweb", 1.0774, 0.0011, 0.0000, 0.0011, 1.0376, -0.0398, -3.70, -36.4), ("paloma_c4_en", 1.0196, 0.0011, 0.0000, 0.0011, 0.9847, -0.0349, -3.43, -30.8), ("paloma_c4_100_domains", 0.9690, 0.0015, 0.0000, 0.0015, 0.9418, -0.0272, -2.80, -18.6), ("paloma_wikitext_103", 0.8665, 0.0023, 0.0000, 0.0023, 0.8957, +0.0292, +3.37, +12.6), ("paloma_redpajama", 1.0265, 0.0040, 0.0000, 0.0040, 1.0603, +0.0338, +3.29, +8.4), ("paloma_dolma-v1_5", 0.9828, 0.0020, 0.0000, 0.0020, 0.9959, +0.0130, +1.33, +6.4), ("paloma_m2d2_wikipedia_unsplit", 0.8894, 0.0020, 0.0000, 0.0020, 0.8993, +0.0099, +1.12, +4.9), ("paloma_mc4", 1.0011, 0.0018, 0.0000, 0.0018, 1.0068, +0.0057, +0.57, +3.2), ("paloma_dolma_100_subreddits", 1.1223, 0.0009, 0.0000, 0.0009, 1.1199, -0.0024, -0.21, -2.7), ("paloma_m2d2_s2orc_unsplit", 0.9160, 0.0018, 0.0000, 0.0018, 0.9189, +0.0029, +0.32, +1.6), ] HQ_V2 = [ ("lambada_openai", 39.6, 0.311, 0.000, 0.686, 0.311, 43.5, +3.9, +9.92, +12.6), ("arc_challenge", 30.9, 0.215, 0.000, 1.356, 0.215, 31.9, +1.0, +3.22, +4.6), ("sciq", 79.0, 0.764, 0.000, 1.258, 0.764, 81.3, +2.3, +2.87, +3.0), ("winogrande", 53.4, 0.721, 0.000, 1.403, 0.721, 51.9, -1.6, -2.94, -2.2), ("logiqa", 21.1, 0.900, 0.000, 1.624, 0.900, 23.0, +1.9, +9.22, +2.2), ("triviaqa", 5.6, 1.525, 0.000, 0.189, 1.525, 8.3, +2.8, +49.38, +1.8), ("social_iqa", 40.5, 0.795, 0.000, 1.113, 0.795, 41.9, +1.3, +3.24, +1.6), ("commonsense_qa", 19.8, 0.855, 0.000, 1.157, 0.855, 21.2, +1.4, +7.02, +1.6), ("boolq", 57.4, 1.647, 0.000, 0.861, 1.647, 59.9, +2.6, +4.48, +1.6), ("copa", 72.3, 1.528, 0.000, 4.546, 1.528, 70.0, -2.3, -3.23, -1.5), ("anli_r2", 33.8, 0.379, 0.000, 1.500, 0.379, 34.4, +0.6, +1.67, +1.5), ("blimp", 81.3, 0.894, 0.002, 0.135, 0.894, 82.5, +1.2, +1.46, +1.3), ("nq_open", 3.2, 0.215, 0.000, 0.300, 0.215, 3.5, +0.2, +7.23, +1.1), ("anli_r1", 31.8, 1.480, 0.000, 1.481, 1.480, 33.2, +1.4, +4.40, +0.9), ("anli_r3", 34.8, 2.014, 0.000, 1.368, 2.014, 33.1, -1.7, -4.95, -0.8), ("arc_easy", 59.6, 0.973, 0.000, 1.004, 0.973, 60.4, +0.7, +1.25, +0.8), ("pubmedqa", 54.2, 1.186, 0.000, 2.228, 1.186, 55.0, +0.8, +1.56, +0.7), ("rte", 49.5, 1.532, 0.000, 3.008, 1.532, 48.4, -1.1, -2.19, -0.7), ("openbookqa", 34.6, 0.907, 0.082, 2.134, 0.911, 35.2, +0.6, +1.64, +0.6), ("squadv2", 256.9, 25.319, 8.912, None, 26.842, 269.5, +12.7, +4.93, +0.5), ("piqa", 70.2, 0.649, 0.000, 1.066, 0.649, 70.4, +0.2, +0.32, +0.3), ("race", 31.3, 0.829, 0.000, 1.436, 0.829, 31.4, +0.1, +0.31, +0.1), ("cb", 49.4, 11.619, 0.000, 6.660, 11.619, 48.2, -1.2, -2.41, -0.1), ("mmlu", 25.2, 0.716, 0.000, 0.366, 0.716, 25.2, +0.1, +0.22, +0.1), ("hellaswag", 49.3, 0.398, 0.000, 0.499, 0.398, 49.3, -0.0, -0.05, -0.1), ("truthfulqa_mc2", 37.9, 0.704, 0.000, 1.395, 0.704, 38.0, +0.0, +0.11, +0.1), ("coqa", 23.2, 0.854, 1.269, 1.616, 1.530, 23.2, -0.0, -0.06, -0.0), ] HQ_V2_PALOMA = [ ("paloma_falcon-refinedweb", 1.0774, 0.0011, 0.0000, 0.0011, 1.0077, -0.0697, -6.47, -63.6), ("paloma_c4_en", 1.0196, 0.0011, 0.0000, 0.0011, 0.9670, -0.0526, -5.16, -46.3), ("paloma_dolma_100_subreddits", 1.1223, 0.0009, 0.0000, 0.0009, 1.0895, -0.0327, -2.92, -36.9), ("paloma_c4_100_domains", 0.9690, 0.0015, 0.0000, 0.0015, 0.9253, -0.0437, -4.51, -30.0), ("paloma_dolma-v1_5", 0.9828, 0.0020, 0.0000, 0.0020, 0.9616, -0.0213, -2.16, -10.5), ("paloma_redpajama", 1.0265, 0.0040, 0.0000, 0.0040, 1.0062, -0.0203, -1.98, -5.0), ("paloma_m2d2_s2orc_unsplit", 0.9160, 0.0018, 0.0000, 0.0018, 0.9245, +0.0085, +0.93, +4.8), ("paloma_wikitext_103", 0.8665, 0.0023, 0.0000, 0.0023, 0.8728, +0.0063, +0.73, +2.7), ("paloma_ptb", 1.0378, 0.0012, 0.0000, 0.0012, 1.0364, -0.0014, -0.13, -1.1), ("paloma_mc4", 1.0011, 0.0018, 0.0000, 0.0018, 1.0021, +0.0010, +0.10, +0.6), ("paloma_m2d2_wikipedia_unsplit", 0.8894, 0.0020, 0.0000, 0.0020, 0.8894, +0.0000, +0.00, +0.0), ] # -------------------------------------------------------------------------- # MQ tier # -------------------------------------------------------------------------- MQ_V1 = [ ("anli_r2", 33.37, 0.115, 0.000, 1.478, 0.115, 31.70, -1.67, -5.00, -14.43), ("lambada_openai", 45.23, 0.465, 0.000, 0.695, 0.465, 49.40, +4.17, +9.22, +8.96), ("social_iqa", 40.69, 0.000, 0.021, 1.112, 0.021, 40.86, +0.17, +0.42, +8.16), ("hellaswag", 50.54, 0.201, 0.000, 0.499, 0.201, 49.05, -1.49, -2.95, -7.41), ("boolq", 57.37, 0.250, 0.000, 0.862, 0.250, 58.56, +1.19, +2.08, +4.76), ("mmlu", 24.20, 0.291, 0.000, 0.359, 0.291, 23.46, -0.74, -3.06, -2.55), ("sciq", 72.20, 0.624, 0.058, 1.408, 0.627, 73.27, +1.07, +1.48, +1.70), ("openbookqa", 31.00, 1.114, 0.000, 2.091, 1.114, 32.60, +1.60, +5.16, +1.44), ("truthfulqa_mc2", 37.86, 0.999, 0.000, 1.395, 0.999, 39.13, +1.27, +3.35, +1.27), ("logiqa", 19.87, 0.582, 0.000, 1.572, 0.582, 20.58, +0.72, +3.61, +1.23), ("squadv2", 251.22, 16.694, 5.189, None, 17.482, 231.56, -19.65, -7.82, -1.12), ("arc_easy", 47.42, 0.947, 0.000, 1.024, 0.947, 46.40, -1.02, -2.14, -1.07), ("pubmedqa", 57.73, 2.603, 0.000, 2.202, 2.603, 60.30, +2.57, +4.45, +0.99), ("coqa", 25.95, 1.609, 0.825, None, 1.809, 24.35, -1.60, -6.18, -0.89), ("piqa", 71.62, 0.635, 0.022, 1.055, 0.636, 71.13, -0.49, -0.68, -0.77), ("copa", 70.33, 4.163, 0.000, 4.454, 4.163, 73.50, +3.17, +4.50, +0.76), ("blimp", 82.04, 0.387, 0.002, 0.133, 0.387, 81.76, -0.27, -0.33, -0.71), ("arc_challenge", 26.17, 0.549, 0.000, 1.277, 0.549, 25.81, -0.36, -1.36, -0.65), ("rte", 55.35, 1.988, 0.000, 2.995, 1.988, 54.15, -1.20, -2.17, -0.61), ("commonsense_qa", 20.42, 0.557, 0.000, 1.160, 0.557, 20.72, +0.30, +1.47, +0.54), ("cb", 26.19, 12.918, 0.000, 6.120, 12.918, 32.14, +5.95, +22.73, +0.46), ("anli_r1", 32.57, 1.137, 0.000, 1.479, 1.137, 32.05, -0.52, -1.59, -0.45), ("nq_open", 1.99, 0.363, 0.000, None, 0.363, 1.88, -0.11, -5.56, -0.30), ("race", 33.59, 1.199, 0.000, 1.461, 1.199, 33.92, +0.33, +1.00, +0.28), ("anli_r3", 34.69, 0.747, 0.000, 1.373, 0.747, 34.88, +0.18, +0.52, +0.24), ("triviaqa", 3.53, 0.694, 0.000, None, 0.694, 3.65, +0.12, +3.34, +0.17), ("winogrande", 54.82, 1.096, 0.032, 1.398, 1.097, 54.70, -0.12, -0.22, -0.11), ] MQ_V1_PALOMA = [ ("paloma_c4_100_domains", 0.9125, 0.0007, 0.0000, 0.0007, 0.8797, -0.0327, -3.59, -47.58), ("paloma_redpajama", 1.0273, 0.0010, 0.0000, 0.0010, 1.0620, +0.0347, +3.38, +33.80), ("paloma_c4_en", 0.9766, 0.0012, 0.0000, 0.0012, 0.9359, -0.0407, -4.17, -33.45), ("paloma_falcon-refinedweb", 1.0274, 0.0015, 0.0000, 0.0015, 0.9787, -0.0487, -4.74, -33.21), ("paloma_ptb", 0.9813, 0.0086, 0.0000, 0.0086, 1.0581, +0.0768, +7.82, +8.94), ("paloma_mc4", 0.9661, 0.0009, 0.0000, 0.0009, 0.9727, +0.0066, +0.68, +7.74), ("paloma_dolma-v1_5", 0.9695, 0.0010, 0.0000, 0.0010, 0.9761, +0.0066, +0.68, +6.60), ("paloma_wikitext_103", 0.8489, 0.0047, 0.0000, 0.0047, 0.8720, +0.0231, +2.73, +4.93), ("paloma_m2d2_wikipedia_unsplit", 0.8740, 0.0020, 0.0000, 0.0020, 0.8812, +0.0072, +0.82, +3.65), ("paloma_m2d2_s2orc_unsplit", 0.9145, 0.0022, 0.0000, 0.0022, 0.9179, +0.0034, +0.37, +1.57), ("paloma_dolma_100_subreddits", 1.1102, 0.0013, 0.0000, 0.0013, 1.1117, +0.0015, +0.14, +1.16), ] MQ_V2 = [ ("lambada_openai", 45.23, 0.465, 0.000, 0.695, 0.465, 49.31, +4.08, +9.02, +8.78), ("anli_r2", 33.37, 0.115, 0.000, 1.497, 0.115, 34.20, +0.83, +2.50, +7.22), ("arc_challenge", 26.17, 0.549, 0.000, 1.265, 0.549, 23.98, -2.19, -8.37, -3.99), ("sciq", 72.20, 0.624, 0.000, 1.402, 0.624, 74.10, +1.90, +2.63, +3.04), ("hellaswag", 50.54, 0.201, 0.000, 0.499, 0.201, 50.09, -0.45, -0.89, -2.24), ("coqa", 25.95, 1.609, 0.000, None, 1.609, 22.63, -3.33, -12.82, -2.07), ("blimp", 82.04, 0.387, 0.002, 0.134, 0.387, 81.26, -0.77, -0.94, -2.00), ("cb", 26.19, 12.918, 0.000, 6.345, 12.918, 50.00, +23.81, +90.91, +1.84), ("truthfulqa_mc2", 37.86, 0.999, 0.000, 1.402, 0.999, 39.47, +1.61, +4.25, +1.61), ("logiqa", 19.87, 0.582, 0.000, 1.554, 0.582, 19.05, -0.82, -4.12, -1.41), ("piqa", 71.62, 0.635, 0.000, 1.057, 0.635, 70.78, -0.83, -1.16, -1.31), ("anli_r1", 32.57, 1.137, 0.000, 1.493, 1.137, 34.00, +1.43, +4.40, +1.26), ("arc_easy", 47.42, 0.947, 0.000, 1.024, 0.947, 46.34, -1.08, -2.28, -1.14), ("boolq", 57.37, 0.250, 0.000, 0.864, 0.250, 57.65, +0.28, +0.48, +1.10), ("openbookqa", 31.00, 1.114, 0.000, 2.084, 1.114, 32.20, +1.20, +3.87, +1.08), ("triviaqa", 3.53, 0.694, 0.000, None, 0.694, 2.83, -0.70, -19.79, -1.01), ("race", 33.59, 1.199, 0.000, 1.453, 1.199, 32.54, -1.05, -3.13, -0.88), ("pubmedqa", 57.73, 2.603, 0.000, 2.202, 2.603, 60.00, +2.27, +3.93, +0.87), ("squadv2", 251.22, 16.694, 6.629, None, 17.963, 266.53, +15.31, +6.10, +0.85), ("anli_r3", 34.69, 0.747, 0.000, 1.376, 0.747, 35.25, +0.56, +1.60, +0.74), ("social_iqa", 40.69, 0.000, 0.000, 1.110, 1.110, 39.87, -0.82, -2.01, -0.74), ("winogrande", 54.82, 1.096, 0.046, 1.399, 1.097, 54.14, -0.68, -1.23, -0.62), ("copa", 70.33, 4.163, 0.000, 4.587, 4.163, 68.00, -2.33, -3.32, -0.56), ("mmlu", 24.20, 0.291, 0.000, 0.361, 0.291, 24.04, -0.16, -0.67, -0.55), ("commonsense_qa", 20.42, 0.557, 0.000, 1.154, 0.557, 20.15, -0.27, -1.34, -0.49), ("nq_open", 1.99, 0.363, 0.000, None, 0.363, 1.86, -0.14, -6.94, -0.38), ("rte", 55.35, 1.988, 0.000, 2.992, 1.988, 55.60, +0.24, +0.43, +0.12), ] MQ_V2_PALOMA = [ ("paloma_c4_100_domains", 0.9125, 0.0007, 0.0000, 0.0007, 0.8721, -0.0403, -4.42, -58.65), ("paloma_falcon-refinedweb", 1.0274, 0.0015, 0.0000, 0.0015, 0.9622, -0.0652, -6.35, -44.51), ("paloma_c4_en", 0.9766, 0.0012, 0.0000, 0.0012, 0.9271, -0.0495, -5.07, -40.64), ("paloma_dolma_100_subreddits", 1.1102, 0.0013, 0.0000, 0.0013, 1.0813, -0.0289, -2.60, -21.76), ("paloma_redpajama", 1.0273, 0.0010, 0.0000, 0.0010, 1.0076, -0.0197, -1.92, -19.18), ("paloma_dolma-v1_5", 0.9695, 0.0010, 0.0000, 0.0010, 0.9507, -0.0188, -1.93, -18.70), ("paloma_m2d2_s2orc_unsplit", 0.9145, 0.0022, 0.0000, 0.0022, 0.9252, +0.0107, +1.17, +4.91), ("paloma_wikitext_103", 0.8489, 0.0047, 0.0000, 0.0047, 0.8704, +0.0215, +2.54, +4.59), ("paloma_mc4", 0.9661, 0.0009, 0.0000, 0.0009, 0.9698, +0.0037, +0.39, +4.39), ("paloma_ptb", 0.9813, 0.0086, 0.0000, 0.0086, 0.9765, -0.0047, -0.48, -0.55), ("paloma_m2d2_wikipedia_unsplit", 0.8740, 0.0020, 0.0000, 0.0020, 0.8737, -0.0003, -0.03, -0.14), ] # -------------------------------------------------------------------------- # LQ tier # -------------------------------------------------------------------------- LQ_V1 = [ ("lambada_openai", 35.7, 0.256, 0.000, 0.673, 0.256, 38.6, +2.9, +7.99, +11.1), ("mmlu", 23.3, 0.144, 0.000, 0.355, 0.144, 22.9, -0.4, -1.60, -2.6), ("anli_r3", 35.3, 0.966, 0.000, 1.373, 0.966, 33.5, -1.8, -5.04, -1.8), ("logiqa", 21.7, 0.582, 0.063, 1.620, 0.585, 22.8, +1.1, +4.87, +1.8), ("copa", 68.0, 1.000, 0.000, 4.662, 1.000, 69.7, +1.7, +2.45, +1.7), ("piqa", 69.2, 0.196, 0.000, 1.075, 0.196, 69.5, +0.3, +0.45, +1.6), ("rte", 52.9, 0.255, 0.000, 3.004, 0.255, 53.2, +0.4, +0.68, +1.4), ("arc_easy", 35.3, 0.611, 0.000, 0.983, 0.611, 36.1, +0.8, +2.33, +1.3), ("hellaswag", 44.0, 0.318, 0.000, 0.495, 0.318, 43.6, -0.4, -0.90, -1.2), ("cb", 41.7, 4.494, 0.000, 6.665, 4.494, 47.0, +5.4, +12.86, +1.2), ("arc_challenge", 23.1, 0.421, 0.000, 1.229, 0.421, 22.7, -0.5, -2.09, -1.1), ("anli_r1", 33.2, 0.819, 0.000, 1.482, 0.819, 32.3, -0.9, -2.61, -1.1), ("openbookqa", 27.6, 0.872, 0.000, 1.990, 0.872, 26.7, -0.9, -3.14, -1.0), ("anli_r2", 33.4, 0.551, 0.000, 1.488, 0.551, 33.0, -0.4, -1.20, -0.7), ("nq_open", 0.8, 0.039, 0.000, 0.147, 0.039, 0.8, +0.0, +3.57, +0.7), ("blimp", 78.4, 0.885, 0.000, 0.141, 0.885, 78.9, +0.6, +0.70, +0.6), ("coqa", 15.6, 1.072, 0.667, 1.439, 1.263, 16.4, +0.8, +4.93, +0.6), ("pubmedqa", 49.5, 5.163, 0.000, 2.234, 5.163, 52.6, +3.1, +6.19, +0.6), ("triviaqa", 1.4, 0.753, 0.000, 0.080, 0.753, 1.1, -0.3, -22.38, -0.4), ("winogrande", 53.2, 1.539, 0.000, 1.402, 1.539, 52.9, -0.4, -0.74, -0.3), ("race", 29.3, 0.722, 0.000, 1.410, 0.722, 29.4, +0.2, +0.54, +0.2), ("boolq", 55.7, 4.145, 0.000, 0.866, 4.145, 56.3, +0.6, +1.12, +0.1), ("commonsense_qa", 19.6, 0.189, 0.000, 1.136, 0.189, 19.6, +0.0, +0.14, +0.1), ("squadv2", 173.3, 24.523, 8.208, None, 25.860, 170.7, -2.6, -1.52, -0.1), ("truthfulqa_mc2", 40.9, 0.674, 0.000, 1.469, 0.674, 40.9, -0.1, -0.16, -0.1), ("social_iqa", 38.8, 0.581, 0.021, 1.102, 0.582, 38.8, +0.0, +0.10, +0.1), ("sciq", 65.6, 2.248, 0.000, 1.494, 2.248, 65.7, +0.1, +0.20, +0.1), ] LQ_V1_PALOMA = [ ("paloma_falcon-refinedweb", 1.0887, 0.0011, 0.0000, 0.0011, 1.0339, -0.0549, -5.04, -51.1), ("paloma_c4_en", 1.0350, 0.0009, 0.0000, 0.0009, 0.9899, -0.0451, -4.36, -47.5), ("paloma_c4_100_domains", 0.9835, 0.0008, 0.0000, 0.0008, 0.9457, -0.0378, -3.84, -46.8), ("paloma_wikitext_103", 0.9694, 0.0010, 0.0000, 0.0010, 1.0033, +0.0340, +3.51, +35.3), ("paloma_ptb", 1.0552, 0.0053, 0.0000, 0.0053, 1.1654, +0.1102, +10.44, +20.9), ("paloma_redpajama", 1.1764, 0.0020, 0.0000, 0.0020, 1.2117, +0.0353, +3.00, +18.0), ("paloma_m2d2_wikipedia_unsplit", 0.9952, 0.0005, 0.0000, 0.0005, 1.0023, +0.0071, +0.71, +15.8), ("paloma_mc4", 1.0309, 0.0008, 0.0000, 0.0008, 1.0354, +0.0045, +0.44, +5.4), ("paloma_dolma-v1_5", 1.1063, 0.0011, 0.0000, 0.0011, 1.1122, +0.0059, +0.54, +5.2), ("paloma_dolma_100_subreddits", 1.1786, 0.0010, 0.0000, 0.0010, 1.1735, -0.0050, -0.43, -4.8), ("paloma_m2d2_s2orc_unsplit", 1.1722, 0.0038, 0.0000, 0.0038, 1.1749, +0.0027, +0.23, +0.7), ] LQ_V2 = [ ("lambada_openai", 35.7, 0.256, 0.000, 0.676, 0.256, 41.0, +5.3, +14.74, +20.6), ("nq_open", 0.8, 0.039, 0.000, 0.121, 0.039, 0.3, -0.4, -57.14, -11.3), ("copa", 68.0, 1.000, 0.000, 4.579, 1.000, 73.0, +5.0, +7.35, +5.0), ("anli_r2", 33.4, 0.551, 0.000, 1.503, 0.551, 35.1, +1.7, +4.99, +3.0), ("mmlu", 23.3, 0.144, 0.000, 0.355, 0.144, 22.9, -0.4, -1.63, -2.6), ("logiqa", 21.7, 0.582, 0.000, 1.631, 0.582, 23.0, +1.3, +6.13, +2.3), ("coqa", 15.6, 1.072, 0.975, 1.464, 1.449, 18.7, +3.1, +19.89, +2.1), ("rte", 52.9, 0.255, 0.000, 3.005, 0.255, 52.4, -0.5, -1.02, -2.1), ("piqa", 69.2, 0.196, 0.000, 1.075, 0.196, 69.6, +0.4, +0.55, +1.9), ("openbookqa", 27.6, 0.872, 0.000, 2.020, 0.872, 29.0, +1.4, +5.07, +1.6), ("cb", 41.7, 4.494, 0.000, 6.683, 4.494, 48.2, +6.5, +15.71, +1.5), ("anli_r3", 35.3, 0.966, 0.000, 1.387, 0.966, 36.7, +1.4, +3.94, +1.4), ("arc_easy", 35.3, 0.611, 0.000, 0.982, 0.611, 35.9, +0.6, +1.79, +1.0), ("winogrande", 53.2, 1.539, 0.032, 1.403, 1.540, 52.0, -1.3, -2.37, -0.8), ("social_iqa", 38.8, 0.581, 0.000, 1.102, 0.581, 38.4, -0.4, -0.92, -0.6), ("boolq", 55.7, 4.145, 0.000, 0.865, 4.145, 57.8, +2.1, +3.79, +0.5), ("truthfulqa_mc2", 40.9, 0.674, 0.000, 1.470, 0.674, 41.2, +0.3, +0.81, +0.5), ("blimp", 78.4, 0.885, 0.000, 0.141, 0.885, 78.7, +0.4, +0.49, +0.4), ("squadv2", 173.3, 24.523, 7.347, None, 25.600, 164.8, -8.6, -4.93, -0.3), ("race", 29.3, 0.722, 0.000, 1.407, 0.722, 29.1, -0.2, -0.65, -0.3), ("anli_r1", 33.2, 0.819, 0.000, 1.489, 0.819, 33.4, +0.2, +0.60, +0.2), ("commonsense_qa", 19.6, 0.189, 0.000, 1.136, 0.189, 19.6, +0.0, +0.14, +0.1), ("arc_challenge", 23.1, 0.421, 0.000, 1.235, 0.421, 23.2, +0.1, +0.25, +0.1), ("pubmedqa", 49.5, 5.163, 0.000, 2.236, 5.163, 50.2, +0.7, +1.35, +0.1), ("sciq", 65.6, 2.248, 0.000, 1.498, 2.248, 65.7, +0.1, +0.20, +0.1), ("hellaswag", 44.0, 0.318, 0.000, 0.495, 0.318, 44.0, +0.0, +0.03, +0.0), ("triviaqa", 1.4, 0.753, 0.000, 0.086, 0.753, 1.4, +0.0, +1.44, +0.0), ] LQ_V2_PALOMA = [ ("paloma_falcon-refinedweb", 1.0887, 0.0011, 0.0000, 0.0011, 1.0171, -0.0717, -6.58, -66.8), ("paloma_c4_en", 1.0350, 0.0009, 0.0000, 0.0009, 0.9808, -0.0542, -5.24, -57.1), ("paloma_c4_100_domains", 0.9835, 0.0008, 0.0000, 0.0008, 0.9375, -0.0460, -4.68, -57.0), ("paloma_dolma_100_subreddits", 1.1786, 0.0010, 0.0000, 0.0010, 1.1481, -0.0304, -2.58, -29.4), ("paloma_dolma-v1_5", 1.1063, 0.0011, 0.0000, 0.0011, 1.0812, -0.0251, -2.27, -22.2), ("paloma_m2d2_wikipedia_unsplit", 0.9952, 0.0005, 0.0000, 0.0005, 0.9897, -0.0055, -0.55, -12.2), ("paloma_redpajama", 1.1764, 0.0020, 0.0000, 0.0020, 1.1529, -0.0235, -2.00, -12.0), ("paloma_wikitext_103", 0.9694, 0.0010, 0.0000, 0.0010, 0.9774, +0.0081, +0.83, +8.4), ("paloma_ptb", 1.0552, 0.0053, 0.0000, 0.0053, 1.0620, +0.0068, +0.64, +1.3), ("paloma_m2d2_s2orc_unsplit", 1.1722, 0.0038, 0.0000, 0.0038, 1.1741, +0.0020, +0.17, +0.5), ("paloma_mc4", 1.0309, 0.0008, 0.0000, 0.0008, 1.0309, -0.0000, -0.00, -0.0), ] RESULTS = { ("HQ", "V1"): (HQ_V1, HQ_V1_PALOMA), ("HQ", "V2"): (HQ_V2, HQ_V2_PALOMA), ("MQ", "V1"): (MQ_V1, MQ_V1_PALOMA), ("MQ", "V2"): (MQ_V2, MQ_V2_PALOMA), ("LQ", "V1"): (LQ_V1, LQ_V1_PALOMA), ("LQ", "V2"): (LQ_V2, LQ_V2_PALOMA), } # Corpus statistics. TIER LABELS: the source table labels the 80.4M-document rows HQ and the # 55.2M-document rows LQ; the run manifests label them the other way round # (LQ = 80.4M, HQ = 55.2M). SWAP_HQ_LQ_LABELS = True applies the manifest labels. # Nothing in the main text depends on the labelling. SWAP_HQ_LQ_LABELS = True CORPUS_STATS = [ ("HQ-Raw", 80_414_180, 34.480), ("HQ-V1", 80_772_480, 34.480), ("HQ-V2", 80_610_629, 34.532), ("MQ-Raw", 39_576_929, 34.440), ("MQ-V1", 39_836_563, 34.440), ("MQ-V2", 40_544_150, 34.489), ("LQ-Raw", 55_207_716, 34.460), ("LQ-V1", 55_619_461, 34.460), ("LQ-V2", 56_051_756, 34.483), ] if SWAP_HQ_LQ_LABELS: CORPUS_STATS = [(g.replace("HQ", "@@").replace("LQ", "HQ").replace("@@", "LQ"), d, t) for g, d, t in CORPUS_STATS] CORPUS_STATS = sorted(CORPUS_STATS, key=lambda r: ({"HQ": 0, "MQ": 1, "LQ": 2}[r[0][:2]], {"Raw": 0, "V1": 1, "V2": 2}[r[0][3:]])) # Recovery yield: share of documents on which the V2 # pipeline performed at least one structural/linguistic repair. RECOVERY_YIELD = {"HQ": 72.63, "MQ": 57.82, "LQ": 56.52} # Repair-operation distribution, in % REPAIR_OPS = { "HQ": {"ftfy": 15.70, "grammar_syntax": 12.02, "char_noise": 5.69, "artifact": 9.45, "line": 5.67, "ocr": 0.13}, "MQ": {"ftfy": 26.09, "grammar_syntax": 26.38, "char_noise": 7.96, "artifact": 15.76, "line": 15.48, "ocr": 0.34}, "LQ": {"ftfy": 17.61, "grammar_syntax": 37.23, "char_noise": 11.27, "artifact": 16.46, "line": 12.69, "ocr": 0.36}, } # Document-level quality shift on 10,000 sampled V2 documents (values read from the # annotations of the quality-shift figures) QUALITY_SHIFT = { "HQ": {"scorer": "DCLM fastText", "pos_share": 97.4, "mean_delta": 0.052641}, "MQ": {"scorer": "DCLM fastText", "pos_share": 95.0, "mean_delta": 0.005511}, "LQ": {"scorer": "Gemini-as-Judge (per the figure captions; the text names DCLM fastText)", "pos_share": 95.0, "mean_delta": 6.2991, "raw_mean": 8.92, "v2_mean": 15.22}, } def verdict(z, thr=2.0, weak=1.0): """Return 'W', 'L', 'w', 'l' or '-' following the evaluation protocol.""" if z is None: return "-" if abs(z) >= thr: return "W" if z > 0 else "L" if abs(z) >= weak: return "w" if z > 0 else "l" return "-" def paloma_verdict(z, thr=2.0, weak=1.0): """For BPB lower is better: negative z is a win.""" if z is None: return "-" if abs(z) >= thr: return "W" if z < 0 else "L" if abs(z) >= weak: return "w" if z < 0 else "l" return "-" def tally(tier, cfg): ds, pal = RESULTS[(tier, cfg)] dW = sum(1 for r in ds if verdict(r[9]) == "W") dL = sum(1 for r in ds if verdict(r[9]) == "L") pW = sum(1 for r in pal if paloma_verdict(r[8]) == "W") pL = sum(1 for r in pal if paloma_verdict(r[8]) == "L") return dict(dW=dW, dL=dL, pW=pW, pL=pL, W=dW + pW, L=dL + pL, net=dW + pW - dL - pL) if __name__ == "__main__": for tier in ("HQ", "MQ", "LQ"): for cfg in ("V1", "V2"): t = tally(tier, cfg) ds, pal = RESULTS[(tier, cfg)] assert len(ds) == 27, (tier, cfg, len(ds)) assert len(pal) == 11, (tier, cfg, len(pal)) print(tier, cfg, t)