[ { "name": { "name": "num_references", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_trials", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_prompt_tokens", "split": "test" }, "count": 1, "sum": 731.5833333333334, "sum_squared": 535214.1736111111, "min": 731.5833333333334, "max": 731.5833333333334, "mean": 731.5833333333334, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_completion_tokens", "split": "test" }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_output_tokens", "split": "test" }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "training_co2_cost", "split": "test" }, "count": 0, "sum": 0, "sum_squared": 0 }, { "name": { "name": "training_energy_cost", "split": "test" }, "count": 0, "sum": 0, "sum_squared": 0 }, { "name": { "name": "inference_runtime", "split": "test" }, "count": 1, "sum": 2.6472620526949564, "sum_squared": 7.007996375638714, "min": 2.6472620526949564, "max": 2.6472620526949564, "mean": 2.6472620526949564, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "batch_size", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_length", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_stop", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_endoftext", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_unknown", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_instances", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "prompt_truncated", "split": "test" }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob", "split": "test" }, "count": 1, "sum": -29.036855126500328, "sum_squared": 843.1389556373683, "min": -29.036855126500328, "max": -29.036855126500328, "mean": -29.036855126500328, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_perplexity_tokens", "split": "test" }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_bytes", "split": "test" }, "count": 1, "sum": 897.2208333333333, "sum_squared": 805005.223767361, "min": 897.2208333333333, "max": 897.2208333333333, "mean": 897.2208333333333, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "perplexity", "split": "test" }, "count": 1, "sum": 1.1822788930548627, "sum_squared": 1.3977833809630313, "min": 1.1822788930548627, "max": 1.1822788930548627, "mean": 1.1822788930548627, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "bits_per_byte", "split": "test" }, "count": 1, "sum": 0.0466900960584917, "sum_squared": 0.002179965069951182, "min": 0.0466900960584917, "max": 0.0466900960584917, "mean": 0.0466900960584917, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob_per_byte", "split": "test" }, "count": 1, "sum": -0.032363108443016535, "sum_squared": 0.0010473707880944481, "min": -0.032363108443016535, "max": -0.032363108443016535, "mean": -0.032363108443016535, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_references", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_references", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_trials", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_trials", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_prompt_tokens", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 731.5833333333334, "sum_squared": 535214.1736111111, "min": 731.5833333333334, "max": 731.5833333333334, "mean": 731.5833333333334, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_prompt_tokens", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 731.5833333333334, "sum_squared": 535214.1736111111, "min": 731.5833333333334, "max": 731.5833333333334, "mean": 731.5833333333334, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_completion_tokens", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_completion_tokens", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_output_tokens", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_output_tokens", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "inference_runtime", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 2.6472620526949564, "sum_squared": 7.007996375638714, "min": 2.6472620526949564, "max": 2.6472620526949564, "mean": 2.6472620526949564, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "inference_runtime", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 2.6472620526949564, "sum_squared": 7.007996375638714, "min": 2.6472620526949564, "max": 2.6472620526949564, "mean": 2.6472620526949564, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "batch_size", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "batch_size", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_length", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_length", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_stop", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_stop", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_endoftext", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_endoftext", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_unknown", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "finish_reason_unknown", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_instances", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_train_instances", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "prompt_truncated", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "prompt_truncated", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 0.0, "sum_squared": 0.0, "min": 0.0, "max": 0.0, "mean": 0.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": -29.036855126500328, "sum_squared": 843.1389556373683, "min": -29.036855126500328, "max": -29.036855126500328, "mean": -29.036855126500328, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": -29.036855126500328, "sum_squared": 843.1389556373683, "min": -29.036855126500328, "max": -29.036855126500328, "mean": -29.036855126500328, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_perplexity_tokens", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_perplexity_tokens", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 173.4125, "sum_squared": 30071.895156249997, "min": 173.4125, "max": 173.4125, "mean": 173.4125, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_bytes", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 897.2208333333333, "sum_squared": 805005.223767361, "min": 897.2208333333333, "max": 897.2208333333333, "mean": 897.2208333333333, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_bytes", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 897.2208333333333, "sum_squared": 805005.223767361, "min": 897.2208333333333, "max": 897.2208333333333, "mean": 897.2208333333333, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_instances", "split": "test" }, "count": 1, "sum": 240.0, "sum_squared": 57600.0, "min": 240.0, "max": 240.0, "mean": 240.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_faithfulness", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_faithfulness", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_faithfulness", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_relevance", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_relevance", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_relevance", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_coherence", "split": "test" }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_coherence", "split": "test", "perturbation": { "name": "robustness", "robustness": true, "fairness": false, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "annotation_call_center_summarization_coherence", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": 1.0, "sum_squared": 1.0, "min": 1.0, "max": 1.0, "mean": 1.0, "variance": 0.0, "stddev": 0.0 } ]