[ { "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": 1688.0, "sum_squared": 2849344.0, "min": 1688.0, "max": 1688.0, "mean": 1688.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_completion_tokens", "split": "test" }, "count": 1, "sum": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_output_tokens", "split": "test" }, "count": 1, "sum": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 1.9109180927276612, "sum_squared": 3.6516079571139226, "min": 1.9109180927276612, "max": 1.9109180927276612, "mean": 1.9109180927276612, "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": -18.555278973548532, "sum_squared": 344.29837778621226, "min": -18.555278973548532, "max": -18.555278973548532, "mean": -18.555278973548532, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_perplexity_tokens", "split": "test" }, "count": 1, "sum": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_bytes", "split": "test" }, "count": 1, "sum": 850.4, "sum_squared": 723180.1599999999, "min": 850.4, "max": 850.4, "mean": 850.4, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "perplexity", "split": "test" }, "count": 1, "sum": 1.1146165042281866, "sum_squared": 1.242369951497863, "min": 1.1146165042281866, "max": 1.1146165042281866, "mean": 1.1146165042281866, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "bits_per_byte", "split": "test" }, "count": 1, "sum": 0.031478844023341626, "sum_squared": 0.0009909176210458708, "min": 0.031478844023341626, "max": 0.031478844023341626, "mean": 0.031478844023341626, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob_per_byte", "split": "test" }, "count": 1, "sum": -0.021819471982065533, "sum_squared": 0.0004760893575761428, "min": -0.021819471982065533, "max": -0.021819471982065533, "mean": -0.021819471982065533, "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": 1688.0, "sum_squared": 2849344.0, "min": 1688.0, "max": 1688.0, "mean": 1688.0, "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": 1688.0, "sum_squared": 2849344.0, "min": 1688.0, "max": 1688.0, "mean": 1688.0, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 1.9109180927276612, "sum_squared": 3.6516079571139226, "min": 1.9109180927276612, "max": 1.9109180927276612, "mean": 1.9109180927276612, "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": 1.9109180927276612, "sum_squared": 3.6516079571139226, "min": 1.9109180927276612, "max": 1.9109180927276612, "mean": 1.9109180927276612, "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": -18.555278973548532, "sum_squared": 344.29837778621226, "min": -18.555278973548532, "max": -18.555278973548532, "mean": -18.555278973548532, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "logprob", "split": "test", "perturbation": { "name": "fairness", "robustness": false, "fairness": true, "computed_on": "worst" } }, "count": 1, "sum": -18.555278973548532, "sum_squared": 344.29837778621226, "min": -18.555278973548532, "max": -18.555278973548532, "mean": -18.555278973548532, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 171.0, "sum_squared": 29241.0, "min": 171.0, "max": 171.0, "mean": 171.0, "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": 850.4, "sum_squared": 723180.1599999999, "min": 850.4, "max": 850.4, "mean": 850.4, "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": 850.4, "sum_squared": 723180.1599999999, "min": 850.4, "max": 850.4, "mean": 850.4, "variance": 0.0, "stddev": 0.0 }, { "name": { "name": "num_instances", "split": "test" }, "count": 1, "sum": 5.0, "sum_squared": 25.0, "min": 5.0, "max": 5.0, "mean": 5.0, "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 } ]