{ "config_general": { "lighteval_sha": "167773f1d5d1647c60dadc31c9e731ab7dbcbbad", "num_few_shot_default": 0, "num_fewshot_seeds": 1, "override_batch_size": 1, "max_samples": null, "job_id": "", "model_name": "TheBloke/LongChat-13B-GPTQ", "model_sha": "8ec25a29033b7be5daeafa26f08e1ea7cf232b98", "model_dtype": "None", "model_size": "6.92 GB" }, "results": { "harness|drop|3": { "em": 0.0, "em_stderr": 0.0, "f1": 5.243288590604027e-05, "f1_stderr": 5.243288590604094e-05 }, "harness|gsm8k|5": { "acc": 0.0, "acc_stderr": 0.0 }, "harness|winogrande|5": { "acc": 0.5114443567482242, "acc_stderr": 0.014048804199859329 }, "all": { "em": 0.0, "em_stderr": 0.0, "f1": 5.243288590604027e-05, "f1_stderr": 5.243288590604094e-05, "acc": 0.2557221783741121, "acc_stderr": 0.007024402099929664 } }, "versions": { "all": 0, "harness|drop|3": 1, "harness|gsm8k|5": 0, "harness|winogrande|5": 0 }, "config_tasks": { "harness|drop": "LM Harness task", "harness|gsm8k": "LM Harness task", "harness|winogrande": "LM Harness task" }, "summary_tasks": { "harness|drop|3": { "hashes": { "hash_examples": "1d27416e8324e9a3", "hash_full_prompts": "a5513ff9a741b385", "hash_input_tokens": "fbe785c62d941897", "hash_cont_tokens": "e72093e065ea3712" }, "truncated": 0, "non_truncated": 9536, "padded": 0, "non_padded": 9536, "effective_few_shots": 3.0, "num_truncated_few_shots": 0 }, "harness|gsm8k|5": { "hashes": { "hash_examples": "4c0843a5d99bcfdc", "hash_full_prompts": "41d55e83abc0e02d", "hash_input_tokens": "bda342e47b5099b2", "hash_cont_tokens": "a16b4401961bdbd8" }, "truncated": 0, "non_truncated": 1319, "padded": 0, "non_padded": 1319, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 }, "harness|winogrande|5": { "hashes": { "hash_examples": "aada0a176fd81218", "hash_full_prompts": "c8655cbd12de8409", "hash_input_tokens": "c0bedf98cb040854", "hash_cont_tokens": "f08975ad6f2d5864" }, "truncated": 0, "non_truncated": 1267, "padded": 2432, "non_padded": 102, "effective_few_shots": 5.0, "num_truncated_few_shots": 0 } }, "summary_general": { "hashes": { "hash_examples": "9b4d8993161e637d", "hash_full_prompts": "08215e527b7e60a5", "hash_input_tokens": "9f72a6bc6743d18f", "hash_cont_tokens": "c6390464e333ce13" }, "truncated": 0, "non_truncated": 12122, "padded": 2432, "non_padded": 10957, "num_truncated_few_shots": 0, "total_evaluation_time_secondes": 0 } }