Data Attribution Methods Leaderboards
Survey and ranking of data attribution methods on data selection and downstream application tasks for the Date-LM Evaluation paper.
Leaderboard Submission:
- To submit your team's scores, click on the "Submit Scores" tab.
Data Attribution Method Categories:
- Gradient (ex. GradDot, GradSim, LESS, DataInf, EKFAC)
- Similarity (ex. RepSim)
- Modeling (ex. MATES)
- Lexical (ex. BM25)
- Baseline (ex. GradSafe, OpenAI Moderation, LLM Classifiers)
- Other
Search Feature:
- Input the name of the method you would like to search / filter for, and then press "Enter". The original row from the leaderboard table will be displayed.
DATE-LM Task Description: Trained pythia-1B model on Fineweb using Lambada reference dataset. Testing results conducted on 10K step checkpoint.
Ranking Metric: highest score in avg column
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "avg",
- "sciq",
- "arc_easy",
- "arc_challenge",
- "logiqa",
- "boolq",
- "hellaswag",
- "piqa",
- "winogrande",
- "openbookqa",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "Rep Sim",
- "Similarity",
- "Pythia-1b",
- "1B",
- 46,
- 0.691,
- 0.441,
- 0.237,
- 0.275,
- 0.561,
- 0.409,
- 0.695,
- 0.537,
- 0.294,
- ""
- [
- 2,
- "Grad Sim",
- "Gradient",
- "Pythia-1b",
- "1B",
- 45.98,
- 0.689,
- 0.44,
- 0.24,
- 0.272,
- 0.556,
- 0.406,
- 0.69,
- 0.537,
- 0.308,
- ""
- [
- 3,
- "Edu",
- "Other",
- "Pythia-1b",
- "1B",
- 45.83,
- 0.688,
- 0.452,
- 0.24,
- 0.264,
- 0.571,
- 0.409,
- 0.689,
- 0.52,
- 0.292,
- ""
- [
- 4,
- "Mates",
- "Modeling",
- "Pythia-1b",
- "1B",
- 45.76,
- 0.685,
- 0.441,
- 0.241,
- 0.269,
- 0.563,
- 0.408,
- 0.696,
- 0.523,
- 0.292,
- ""
- [
- 5,
- "BM25",
- "Lexical",
- "Pythia-1b",
- "1B",
- 45.72,
- 0.692,
- 0.439,
- 0.239,
- 0.26,
- 0.556,
- 0.406,
- 0.696,
- 0.531,
- 0.296,
- ""
- [
- 6,
- "Random",
- "Other",
- "Pythia-1b",
- "1B",
- 45.34,
- 0.689,
- 0.431,
- 0.244,
- 0.275,
- 0.52,
- 0.407,
- 0.69,
- 0.535,
- 0.29,
- ""
- [
- "metadata": null
DATE-LM Task Description: Trained pythia-1B model on Fineweb using Lambada reference dataset. Testing results conducted on 30K step checkpoint.
Ranking Metric: highest score in avg column
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "avg",
- "sciq",
- "arc_easy",
- "arc_challenge",
- "logiqa",
- "boolq",
- "hellaswag",
- "piqa",
- "winogrande",
- "openbookqa",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "Edu",
- "Other",
- "Pythia-1b",
- "1B",
- 48.02,
- 0.7,
- 0.471,
- 0.262,
- 0.267,
- 0.616,
- 0.474,
- 0.709,
- 0.511,
- 0.312,
- ""
- [
- 2,
- "Grad Sim",
- "Gradient",
- "Pythia-1b",
- "1B",
- 47.92,
- 0.711,
- 0.465,
- 0.254,
- 0.273,
- 0.606,
- 0.475,
- 0.711,
- 0.522,
- 0.296,
- ""
- [
- 3,
- "Rep Sim",
- "Similarity",
- "Pythia-1b",
- "1B",
- 47.83,
- 0.697,
- 0.462,
- 0.259,
- 0.263,
- 0.612,
- 0.474,
- 0.712,
- 0.526,
- 0.3,
- ""
- [
- 4,
- "Mates",
- "Modeling",
- "Pythia-1b",
- "1B",
- 47.64,
- 0.702,
- 0.464,
- 0.253,
- 0.26,
- 0.617,
- 0.474,
- 0.708,
- 0.518,
- 0.292,
- ""
- [
- 5,
- "Random",
- "Other",
- "Pythia-1b",
- "1B",
- 47.49,
- 0.695,
- 0.456,
- 0.255,
- 0.276,
- 0.602,
- 0.475,
- 0.711,
- 0.51,
- 0.294,
- ""
- [
- "metadata": null
DATE-LM Task Description: Targeted instruction tuning setting. Given a diverse instruction set and a eval dataset, we select data that would yield optimal performance on the eval data. For this task, the training data pool is Tulu3 (unfiltered) and the eval data is MMLU, GSM8K, and BBH.
Ranking Metric: average of the MMLU, GSM8K, and BBH scores
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "MMLU",
- "GSM8K",
- "BBH",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "RDS+",
- "Similarity",
- "Llama-3.1-8B",
- "8B",
- 62.4,
- 59.6,
- 66.9,
- ""
- [
- 2,
- "Rep Sim",
- "Similarity",
- "Llama-3.1-8B",
- "8B",
- 61.2,
- 59.2,
- 65.9,
- ""
- [
- 3,
- "Random Avg",
- "Other",
- "Llama-3.1-8B",
- "8B",
- 60.2,
- 59.6,
- 65.6,
- ""
- [
- 4,
- "LESS (optimizer)",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 60,
- 59.5,
- 64.2,
- ""
- [
- 5,
- "BM25",
- "Lexical",
- "Llama-3.1-8B",
- "8B",
- 59.5,
- 60.2,
- 62.5,
- ""
- [
- 6,
- "Grad Sim",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 58.4,
- 57.8,
- 65.5,
- ""
- [
- "metadata": null
DATE-LM Task Description: This leaderboard presents detection AUPRC results of baseline methods and data attribution methods in the homogenous setting (i.e., detecting small amount of toxic/biased data embedded into larger benign data).
Ranking Metric: AUPRC (an average of ToxicChat, XSTest-response, JailBreakBench)
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "ToxicChat",
- "XSTest-response",
- "JailBreakBench",
- "AUPRC",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "Rep-Sim",
- "Similarity",
- "Llama-3.1-8B",
- "8B",
- 0.989,
- 0.999,
- 0.98,
- 0.989,
- ""
- [
- 2,
- "Wildguard",
- "Baseline",
- "N/A",
- "N/A",
- 0.56,
- 0.93,
- 0.989,
- 0.827,
- ""
- [
- 3,
- "Llama-Guard-3-8B",
- "Baseline",
- "N/A",
- "N/A",
- 0.445,
- 0.916,
- 0.985,
- 0.782,
- ""
- [
- 4,
- "Rep-Sim",
- "Similarity",
- "Llama-3.2-1B",
- "1B",
- 0.632,
- 0.792,
- 0.854,
- 0.759,
- ""
- [
- 5,
- "LESS",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.388,
- 0.724,
- 1,
- 0.704,
- ""
- [
- 6,
- "LESS",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.294,
- 0.792,
- 1,
- 0.695,
- ""
- [
- 7,
- "Grad Sim",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.259,
- 0.798,
- 1,
- 0.686,
- ""
- [
- 8,
- "Rep-Sim",
- "Similarity",
- "Pythia-1b",
- "1B",
- 0.374,
- 0.657,
- 0.986,
- 0.672,
- ""
- [
- 9,
- "LESS",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.499,
- 0.615,
- 0.767,
- 0.627,
- ""
- [
- 10,
- "EKFAC",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.264,
- 0.562,
- 1,
- 0.609,
- ""
- [
- 11,
- "Grad Sim",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.106,
- 0.647,
- 1,
- 0.584,
- ""
- [
- 12,
- "EKFAC",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.216,
- 0.497,
- 1,
- 0.571,
- ""
- [
- 13,
- "Grad Sim",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.28,
- 0.603,
- 0.82,
- 0.567,
- ""
- [
- 14,
- "DataInf",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.204,
- 0.487,
- 0.999,
- 0.563,
- ""
- [
- 15,
- "DataInf",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.215,
- 0.442,
- 1,
- 0.552,
- ""
- [
- 16,
- "Grad Dot",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.212,
- 0.437,
- 1,
- 0.55,
- ""
- [
- 17,
- "GradSafe",
- "Baseline",
- "N/A",
- "N/A",
- 0.347,
- 0.491,
- 0.802,
- 0.546,
- ""
- [
- 18,
- "ShieldGemma-2b",
- "Baseline",
- "N/A",
- "N/A",
- 0.17,
- 0.74,
- 0.664,
- 0.525,
- ""
- [
- 19,
- "Grad Dot",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.084,
- 0.483,
- 0.999,
- 0.522,
- ""
- [
- 20,
- "Grad Dot",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.47,
- 0.368,
- 0.274,
- 0.371,
- ""
- [
- 21,
- "AEGIS-Defensive",
- "Baseline",
- "N/A",
- "N/A",
- 0.376,
- 0.274,
- 0.346,
- 0.332,
- ""
- [
- 22,
- "test",
- "Gradient",
- "pythia",
- "1B",
- 0.3,
- 0.3,
- 0.3,
- 0.3,
- "test"
- [
- 23,
- "OpenAI Moderation",
- "Baseline",
- "N/A",
- "N/A",
- 0.243,
- 0.378,
- 0.187,
- 0.269,
- ""
- [
- "metadata": null
DATE-LM Task Description: This leaderboard presents detection AUPRC results of baseline methods and data attribution methods in the heterogeneous setting (i.e., safety-aligned examples that resemble unsafe data in format but contain safe responses).
Ranking Metric: AUPRC (an average of ToxicChat, XSTest-response, JailBreakBench)
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "ToxicChat",
- "XSTest-response",
- "JailBreakBench",
- "AUPRC",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "Wildguard",
- "Baseline",
- "N/A",
- "N/A",
- 0.551,
- 0.928,
- 0.972,
- 0.817,
- ""
- [
- 2,
- "Llama-Guard-3-8B",
- "Baseline",
- "N/A",
- "N/A",
- 0.423,
- 0.91,
- 0.966,
- 0.766,
- ""
- [
- 3,
- "Rep-Sim",
- "Similarity",
- "Llama-3.2-1B",
- "1B",
- 0.598,
- 0.733,
- 0.461,
- 0.597,
- ""
- [
- 4,
- "Rep-Sim",
- "Similarity",
- "Llama-3.1-8B",
- "8B",
- 0.602,
- 0.638,
- 0.514,
- 0.585,
- ""
- [
- 5,
- "GradSafe",
- "Baseline",
- "N/A",
- "N/A",
- 0.347,
- 0.491,
- 0.802,
- 0.546,
- ""
- [
- 6,
- "LESS",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.326,
- 0.734,
- 0.484,
- 0.515,
- ""
- [
- 7,
- "Grad Sim",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.228,
- 0.772,
- 0.531,
- 0.51,
- ""
- [
- 8,
- "Rep-Sim",
- "Similarity",
- "Pythia-1b",
- "1B",
- 0.335,
- 0.58,
- 0.578,
- 0.498,
- ""
- [
- 9,
- "ShieldGemma-2b",
- "Baseline",
- "N/A",
- "N/A",
- 0.165,
- 0.731,
- 0.552,
- 0.483,
- ""
- [
- 10,
- "LESS",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.23,
- 0.616,
- 0.596,
- 0.481,
- ""
- [
- 11,
- "Grad Sim",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.362,
- 0.601,
- 0.434,
- 0.466,
- ""
- [
- 12,
- "Grad Sim",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.223,
- 0.703,
- 0.401,
- 0.442,
- ""
- [
- 13,
- "LESS",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.258,
- 0.744,
- 0.114,
- 0.372,
- ""
- [
- 14,
- "EKFAC",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.239,
- 0.398,
- 0.369,
- 0.334,
- ""
- [
- 15,
- "DataInf",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.195,
- 0.392,
- 0.396,
- 0.328,
- ""
- [
- 16,
- "Grad Dot",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.194,
- 0.389,
- 0.396,
- 0.326,
- ""
- [
- 17,
- "AEGIS-Defensive",
- "Baseline",
- "N/A",
- "N/A",
- 0.376,
- 0.274,
- 0.294,
- 0.314,
- ""
- [
- 18,
- "EKFAC",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.221,
- 0.344,
- 0.373,
- 0.313,
- ""
- [
- 19,
- "DataInf",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.196,
- 0.347,
- 0.369,
- 0.304,
- ""
- [
- 20,
- "Grad Dot",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.195,
- 0.341,
- 0.369,
- 0.302,
- ""
- [
- 21,
- "OpenAI Moderation",
- "Baseline",
- "N/A",
- "N/A",
- 0.214,
- 0.358,
- 0.185,
- 0.253,
- ""
- [
- 22,
- "Grad Dot",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.289,
- 0.328,
- 0.085,
- 0.234,
- ""
- [
- "metadata": null
DATE-LM Task Description: Identifying the specific training examples that support a model's generated facts.
Ranking Metric: average of Recall@50 and MRR
- "headers": [
- "Rank",
- "Method",
- "Attribution Method Type",
- "Model",
- "Model Size",
- "Recall@50",
- "MRR",
- "Paper/Code/Contact Link"
- "data": [
- [
- 1,
- "Grad Sim",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.626,
- 0.97,
- ""
- [
- 2,
- "Rep Sim",
- "Similarity",
- "Llama-3.1-8B",
- "8B",
- 0.625,
- 0.965,
- ""
- [
- 3,
- "LESS",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.491,
- 0.991,
- ""
- [
- 4,
- "Grad Sim",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.584,
- 0.839,
- ""
- [
- 5,
- "LESS",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.573,
- 0.807,
- ""
- [
- 6,
- "EKFAC",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.485,
- 0.881,
- ""
- [
- 7,
- "Grad Sim",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.493,
- 0.836,
- ""
- [
- 8,
- "Rep Sim",
- "Similarity",
- "Llama-3.2-1B",
- "1B",
- 0.552,
- 0.758,
- ""
- [
- 9,
- "LESS",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.5,
- 0.772,
- ""
- [
- 10,
- "DataInf",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.475,
- 0.785,
- ""
- [
- 11,
- "Grad Dot",
- "Gradient",
- "Llama-3.2-1B",
- "1B",
- 0.465,
- 0.786,
- ""
- [
- 12,
- "DataInf",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.472,
- 0.765,
- ""
- [
- 13,
- "Grad Dot",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.466,
- 0.768,
- ""
- [
- 14,
- "EKFAC",
- "Gradient",
- "Pythia-1b",
- "1B",
- 0.465,
- 0.766,
- ""
- [
- 15,
- "Rep Sim",
- "Similarity",
- "Pythia-1b",
- "1B",
- 0.376,
- 0.79,
- ""
- [
- 16,
- "BM25",
- "Lexical",
- "Llama-3.1-8B",
- "8B",
- 0.313,
- 0.826,
- ""
- [
- 17,
- "BM25",
- "Lexical",
- "Pythia-1b",
- "1B",
- 0.305,
- 0.771,
- ""
- [
- 18,
- "BM25",
- "Lexical",
- "Llama-3.2-1B",
- "1B",
- 0.236,
- 0.683,
- ""
- [
- 19,
- "Grad Dot",
- "Gradient",
- "Llama-3.1-8B",
- "8B",
- 0.226,
- 0.303,
- ""
- [
- "metadata": null
Submit Your Score to a Leaderboard
Note: Please first select the leaderboard you would like to submit to. This will display the fields for the corresponding metrics that are needed.