Files
ml_debug/benchmark/results/gpt56-medium/metadata.json
T

54 lines
1.4 KiB
JSON

{
"run_id": "gpt56-medium",
"model": "gpt-5.6-sol",
"workers": 4,
"reasoning_effort": "medium",
"case_ids": [
"case_01",
"case_02",
"case_03",
"case_04",
"case_05",
"case_06",
"case_07",
"case_08",
"case_09",
"case_10"
],
"conditions": [
"control",
"treatment"
],
"skill_sha256": "adbdd538b81357ed43d58f0027386d45b272ffb31ba09a4eab5552ee12030948",
"base_instruction_sha256": "06378fe7d972e6d6f8be0780597132dd9f990ff75a2acac74d59ce6b986233f2",
"fixture_sha256": {
"cases.json": "9c3972377f1ebf635afb522634c0d64c182f6abd53bbb859ea834fca66181307",
"answers.json": "5c27df932406d571c9e06a823b30695d5debd07fd25ed3e51c7ac0143b84fab0",
"response.schema.json": "19bfe551ae38efd4829bc25f53b9b59e956620392fe1f67d737be1a4900db4e6"
},
"codex_version": "codex-cli 0.144.1",
"git_commit": "a0fcfa291e41316924424ba3f2def33eb37b7338",
"job_order": [
"treatment/case_10",
"control/case_03",
"treatment/case_06",
"control/case_01",
"treatment/case_07",
"control/case_09",
"treatment/case_09",
"control/case_02",
"control/case_05",
"treatment/case_02",
"treatment/case_01",
"control/case_10",
"control/case_07",
"treatment/case_04",
"treatment/case_08",
"control/case_08",
"control/case_06",
"treatment/case_05",
"control/case_04",
"treatment/case_03"
]
}