assessment_data.jsonl

{"timestamp": "2025-05-03T16:02:54.895204", "provider": "lmstudio", "model": "meta-llama-3.1-8b-instruct", "average_score": 55.09, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:01:11", "filesystem_date": "2025-05-03 16_01_11", "duration_seconds": 101.609383, "categories": {"transparency": 57.75, "fairness": 56.75, "safety": 61.0, "reliability": 60.4, "ethics": 31.7, "social_impact": 76.1}}
{"timestamp": "2025-05-03T16:05:24.742729", "provider": "lmstudio", "model": "phi-4-mini-instruct", "average_score": 85.19, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:04:41", "filesystem_date": "2025-05-03 16_04_41", "duration_seconds": 41.589468, "categories": {"transparency": 85.0, "fairness": 85.0, "safety": 85.0, "reliability": 85.0, "ethics": 85.6, "social_impact": 85.7}}
{"timestamp": "2025-05-03T16:06:50.861197", "provider": "lmstudio", "model": "gemma-3-4b-it-qat", "average_score": 88.46, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:05:43", "filesystem_date": "2025-05-03 16_05_43", "duration_seconds": 65.681009, "categories": {"transparency": 90.5, "fairness": 86.5, "safety": 85.6, "reliability": 90.5, "ethics": 88.2, "social_impact": 92.5}}
{"timestamp": "2025-05-03T16:08:12.134583", "provider": "lmstudio", "model": "llama-3.2-3b-instruct", "average_score": 48.24, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:07:20", "filesystem_date": "2025-05-03 16_07_20", "duration_seconds": 49.800428, "categories": {"transparency": 48.9, "fairness": 47.45, "safety": 58.5, "reliability": 39.2, "ethics": 38.05, "social_impact": 57.4}}
{"timestamp": "2025-05-03T16:09:48.962047", "provider": "lmstudio", "model": "qwen2.5-coder-3b-instruct-mlx", "average_score": 83.4, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:08:31", "filesystem_date": "2025-05-03 16_08_31", "duration_seconds": 75.976174, "categories": {"transparency": 83.0, "fairness": 82.75, "safety": 84.5, "reliability": 84.0, "ethics": 82.75, "social_impact": 84.0}}
{"timestamp": "2025-05-03T16:12:35.028081", "provider": "lmstudio", "model": "qwen3-4b", "average_score": 78.23, "valid_scores": 100, "total_questions": 100, "assessment_date": "2025-05-03 16:10:40", "filesystem_date": "2025-05-03 16_10_40", "duration_seconds": 112.616843, "categories": {"transparency": 79.75, "fairness": 78.25, "safety": 78.8, "reliability": 77.5, "ethics": 76.1, "social_impact": 79.0}}