MTEB-lite: eng (K=100) and eng-super (K=20) - ids + loader, validation, selection criteria, reference models
3037e25 verified | # MTEB-lite v2 selection report (eng, K=100) | |
| Reference set (28 models, 15 families, <= 3 per family): **A-v3@6750**: A-lin0.5, A-lin0.6, A-ties-d0.1-l0.5; **AS3**: AS3-dareties-p0.9-l1.0, AS3-ties-d0.3-l0.5, AS3-ties-d0.3-l1.0; **AV2**: AV2-dareties-p0.9-l1.0, AV2-lin0.5, AV2-lin1.0; **W**: W-A0.7-S3-0.3-l0.5; **all-MiniLM-L6-v2**: all-MiniLM-L6-v2; **bge-m3**: bge-m3; **e5-base-v2**: e5-base-v2; **embeddinggemma-300m**: embeddinggemma-300m; **multilingual-e5-base**: multilingual-e5-base; **multilingual-e5-small**: multilingual-e5-small; **qwen3-0.6b**: qwen3-0.6b, qwen3-0.6b+norm; **qwen3-4b**: qwen3-4b; **v1-F/P**: F-lin0.5, F-lin1.0, F-ties-d0.1-l0.5; **v4**: S-lin0.5, V-lin0.7, V-ties-d0.1-l0.5; **v5**: S5-lin0.5, V5-ties-d0.1-l0.5, V5-ties-d0.3-l0.7 | |
| K=100 for per-item tasks, K=100 for set-metric tasks. Chosen: **anchor/mean+anchorset** (per-item method/estimator by the family-level LOMO macro tau among {anchor, fisher} x {mean, pirt, gp}; set-metric selection: anchorset). Set-metric tasks (STS, clustering, pair classification): no per-item IRT (the metric is not a sum over items) -> stratified random K ('strat') or unweighted anchor points on per-item outcomes ('anchorset'), + calibration. Random = uniform K, metrics averaged over 10 single draws. tau_res = tau on the pairs whose true score differs by >= 0.5 points. | |
| ## Aggregate, family-level LOMO (20 held-out models with all tasks; scores x100) | |
| | estimate | macro tau | macro tau_res | macro r | macro MAE | mean tau | mean tau_res | mean r | mean MAE | | |
| |---|---|---|---|---|---|---|---|---| | |
| | anchor/mean+strat | 0.848 | 0.919 | 0.991 | 0.429 | 0.881 | 0.938 | 0.993 | 0.541 | | |
| | anchor/mean+anchorset (chosen) | 0.852 | 0.950 | 0.992 | 0.438 | 0.894 | 0.950 | 0.994 | 0.468 | | |
| | anchor/pirt+strat | 0.810 | 0.860 | 0.983 | 0.340 | 0.839 | 0.887 | 0.985 | 0.430 | | |
| | anchor/pirt+anchorset | 0.757 | 0.862 | 0.981 | 0.397 | 0.862 | 0.900 | 0.984 | 0.422 | | |
| | anchor/gp+strat | 0.824 | 0.887 | 0.988 | 0.373 | 0.850 | 0.906 | 0.990 | 0.471 | | |
| | anchor/gp+anchorset | 0.799 | 0.912 | 0.987 | 0.393 | 0.873 | 0.925 | 0.990 | 0.409 | | |
| | fisher/mean+strat | 0.743 | 0.833 | 0.976 | 0.510 | 0.761 | 0.868 | 0.976 | 0.703 | | |
| | fisher/mean+anchorset | 0.725 | 0.849 | 0.978 | 0.512 | 0.820 | 0.887 | 0.980 | 0.566 | | |
| | fisher/pirt+strat | 0.743 | 0.833 | 0.976 | 0.516 | 0.757 | 0.865 | 0.976 | 0.712 | | |
| | fisher/pirt+anchorset | 0.725 | 0.849 | 0.978 | 0.509 | 0.820 | 0.887 | 0.980 | 0.574 | | |
| | fisher/gp+strat | 0.745 | 0.833 | 0.976 | 0.504 | 0.760 | 0.868 | 0.976 | 0.702 | | |
| | fisher/gp+anchorset | 0.725 | 0.849 | 0.978 | 0.502 | 0.820 | 0.887 | 0.980 | 0.564 | | |
| | random | 0.833 | 0.897 | 0.989 | 0.391 | 0.862 | 0.907 | 0.989 | 0.465 | | |
| | pool | 0.892 | 0.937 | 0.994 | 0.168 | 0.860 | 0.912 | 0.995 | 0.211 | | |
| ## Aggregate, distinct-level LOMO (20 held-out models with all tasks; scores x100) | |
| | estimate | macro tau | macro tau_res | macro r | macro MAE | mean tau | mean tau_res | mean r | mean MAE | | |
| |---|---|---|---|---|---|---|---|---| | |
| | anchor/mean+strat | 0.858 | 0.921 | 0.991 | 0.414 | 0.882 | 0.939 | 0.993 | 0.516 | | |
| | anchor/mean+anchorset (chosen) | 0.881 | 0.950 | 0.989 | 0.422 | 0.934 | 0.963 | 0.993 | 0.429 | | |
| | anchor/pirt+strat | 0.815 | 0.862 | 0.983 | 0.337 | 0.844 | 0.889 | 0.985 | 0.417 | | |
| | anchor/pirt+anchorset | 0.807 | 0.862 | 0.979 | 0.420 | 0.871 | 0.887 | 0.984 | 0.415 | | |
| | anchor/gp+strat | 0.835 | 0.887 | 0.987 | 0.363 | 0.862 | 0.908 | 0.990 | 0.453 | | |
| | anchor/gp+anchorset | 0.871 | 0.925 | 0.984 | 0.407 | 0.902 | 0.925 | 0.989 | 0.410 | | |
| | fisher/mean+strat | 0.771 | 0.834 | 0.977 | 0.497 | 0.811 | 0.879 | 0.976 | 0.678 | | |
| | fisher/mean+anchorset | 0.799 | 0.849 | 0.977 | 0.463 | 0.841 | 0.887 | 0.980 | 0.522 | | |
| | fisher/pirt+strat | 0.768 | 0.834 | 0.976 | 0.503 | 0.807 | 0.879 | 0.976 | 0.687 | | |
| | fisher/pirt+anchorset | 0.786 | 0.849 | 0.976 | 0.462 | 0.839 | 0.887 | 0.980 | 0.530 | | |
| | fisher/gp+strat | 0.768 | 0.834 | 0.976 | 0.492 | 0.808 | 0.879 | 0.976 | 0.678 | | |
| | fisher/gp+anchorset | 0.797 | 0.849 | 0.976 | 0.457 | 0.839 | 0.887 | 0.980 | 0.521 | | |
| | random | 0.848 | 0.901 | 0.989 | 0.382 | 0.868 | 0.907 | 0.989 | 0.455 | | |
| | pool | 0.881 | 0.925 | 0.994 | 0.165 | 0.860 | 0.912 | 0.995 | 0.206 | | |
| ## Per task (family-level LOMO; chosen estimator vs random K vs pool) | |
| | task | type | ref | chosen tau | chosen r | chosen MAE | random tau | random MAE | pool tau | pool MAE | | |
| |---|---|---|---|---|---|---|---|---|---| | |
| | ArguAna | Retrieval | 28 | 0.905 | 0.976 | 0.020 | 0.903 | 0.016 | 0.963 | 0.006 | | |
| | ArXivHierarchicalClusteringP2P | Clustering | 28 | 0.650 | 0.956 | 0.009 | 0.650 | 0.009 | 0.629 | 0.007 | | |
| | ArXivHierarchicalClusteringS2S | Clustering | 28 | 0.528 | 0.889 | 0.014 | 0.625 | 0.010 | 0.841 | 0.008 | | |
| | AskUbuntuDupQuestions | Reranking | 27 | 0.726 | 0.987 | 0.005 | 0.628 | 0.007 | 0.960 | 0.000 | | |
| | BIOSSES | STS | 27 | 0.897 | 0.987 | 0.003 | 0.897 | 0.003 | 0.897 | 0.003 | | |
| | Banking77Classification | Classification | 27 | 0.714 | 0.976 | 0.013 | 0.676 | 0.012 | 0.823 | 0.005 | | |
| | BiorxivClusteringP2P.v2 | Clustering | 28 | 0.735 | 0.919 | 0.017 | 0.756 | 0.017 | 0.894 | 0.007 | | |
| | CQADupstackGamingRetrieval | Retrieval | 21 | 0.694 | 0.752 | 0.016 | 0.686 | 0.017 | 0.943 | 0.007 | | |
| | CQADupstackUnixRetrieval | Retrieval | 21 | 0.780 | 0.890 | 0.019 | 0.761 | 0.016 | 0.933 | 0.004 | | |
| | ClimateFEVERHardNegatives | Retrieval | 22 | 0.870 | 0.942 | 0.017 | 0.783 | 0.017 | 0.965 | 0.006 | | |
| | FEVERHardNegatives | Retrieval | 22 | 0.887 | 0.964 | 0.013 | 0.798 | 0.014 | 0.957 | 0.004 | | |
| | FiQA2018 | Retrieval | 27 | 0.829 | 0.988 | 0.012 | 0.742 | 0.014 | 0.909 | 0.007 | | |
| | HotpotQAHardNegatives | Retrieval | 28 | 0.841 | 0.993 | 0.011 | 0.745 | 0.011 | 0.940 | 0.007 | | |
| | ImdbClassification | Classification | 27 | 0.794 | 0.988 | 0.013 | 0.670 | 0.012 | 0.880 | 0.008 | | |
| | MTOPDomainClassification | Classification | 27 | 0.926 | 0.973 | 0.010 | 0.859 | 0.008 | 0.966 | 0.002 | | |
| | MassiveIntentClassification | Classification | 28 | 0.947 | 0.994 | 0.008 | 0.902 | 0.012 | 0.973 | 0.004 | | |
| | MassiveScenarioClassification | Classification | 28 | 0.942 | 0.990 | 0.010 | 0.915 | 0.011 | 0.973 | 0.004 | | |
| | MedrxivClusteringP2P.v2 | Clustering | 28 | 0.618 | 0.631 | 0.023 | 0.643 | 0.014 | 0.804 | 0.005 | | |
| | MedrxivClusteringS2S.v2 | Clustering | 28 | 0.724 | 0.940 | 0.012 | 0.577 | 0.013 | 0.714 | 0.005 | | |
| | MindSmallReranking | Reranking | 27 | 0.291 | -0.553 | 0.016 | 0.341 | 0.008 | 0.869 | 0.002 | | |
| | SCIDOCS | Retrieval | 21 | 0.397 | 0.939 | 0.008 | 0.507 | 0.008 | 0.923 | 0.002 | | |
| | SICK-R | STS | 28 | 0.666 | 0.945 | 0.011 | 0.673 | 0.016 | 0.958 | 0.005 | | |
| | STS12 | STS | 28 | 0.687 | 0.888 | 0.021 | 0.748 | 0.014 | 0.973 | 0.003 | | |
| | STS13 | STS | 28 | 0.942 | 0.986 | 0.014 | 0.892 | 0.011 | 0.973 | 0.002 | | |
| | STS14 | STS | 28 | 0.798 | 0.970 | 0.017 | 0.807 | 0.016 | 0.958 | 0.004 | | |
| | STS15 | STS | 28 | 0.830 | 0.936 | 0.011 | 0.735 | 0.010 | 0.905 | 0.004 | | |
| | STSBenchmark | STS | 28 | 0.867 | 0.953 | 0.014 | 0.877 | 0.010 | 0.973 | 0.002 | | |
| | SprintDuplicateQuestions | PairClassification | 28 | 0.676 | 0.256 | 0.058 | -0.041 | 0.042 | 0.682 | 0.013 | | |
| | StackExchangeClustering.v2 | Clustering | 28 | 0.687 | 0.747 | 0.058 | 0.626 | 0.050 | 0.830 | 0.030 | | |
| | StackExchangeClusteringP2P.v2 | Clustering | 27 | 0.171 | 0.424 | 0.042 | 0.295 | 0.039 | 0.634 | 0.024 | | |
| | TRECCOVID | Retrieval | 22 | 0.487 | 0.919 | 0.033 | 0.487 | 0.033 | 0.635 | 0.024 | | |
| | Touche2020Retrieval.v3 | Retrieval | 20 | 0.757 | 0.904 | 0.018 | 0.757 | 0.018 | 0.757 | 0.017 | | |
| | ToxicConversationsClassification | Classification | 28 | 0.920 | 0.971 | 0.014 | 0.857 | 0.018 | 0.889 | 0.013 | | |
| | TweetSentimentExtractionClassification | Classification | 27 | 0.857 | 0.967 | 0.011 | 0.743 | 0.015 | 0.960 | 0.003 | | |
| | TwentyNewsgroupsClustering.v2 | Clustering | 27 | 0.703 | 0.841 | 0.036 | 0.606 | 0.032 | 0.789 | 0.007 | | |
| | TwitterSemEval2015 | PairClassification | 27 | 0.526 | 0.465 | 0.041 | 0.551 | 0.020 | 0.823 | 0.006 | | |
| | TwitterURLCorpus | PairClassification | 28 | 0.130 | 0.307 | 0.018 | 0.399 | 0.012 | 0.825 | 0.005 | | |
| | SummEvalSummarization.v2 | Summarization | 28 | 0.995 | 0.997 | 0.001 | 0.995 | 0.001 | 0.995 | 0.001 | | |
| | AmazonCounterfactualClassification | Classification | 28 | 0.785 | 0.994 | 0.008 | 0.731 | 0.012 | 0.884 | 0.004 | | |
| | STS17 | STS | 28 | 0.745 | 0.925 | 0.014 | 0.846 | 0.007 | 0.984 | 0.001 | | |
| | STS22.v2 | STS | 22 | 0.878 | 0.912 | 0.005 | 0.713 | 0.009 | 0.974 | 0.003 | | |