20 lines
901 B
JSON
20 lines
901 B
JSON
{
|
|
"method": "Five preselected historical RAG smoke cases; not a new accuracy estimate.",
|
|
"selected_ids": [
|
|
"q0053",
|
|
"q0060",
|
|
"q0608",
|
|
"q1147",
|
|
"q0888"
|
|
],
|
|
"model": "model120-fast",
|
|
"source_file": "tools/rag/evaluation/results/test_classic_rag_frozen/rag_answers_test.json",
|
|
"source_sha256": "eba03639f8a3c32ee47a4e294d8592c1071fc6fd43e38967e39131b840fefa7e",
|
|
"generation_files": {
|
|
"/home/janko/DP/zp-agent/tools/answer_evaluation/check_rag_runtime.py": "37eef74ed8120109ed0c7f2d5a0edba497687a76ab5eaa154bd4626c9edb311c",
|
|
"tools/rag/evaluation/rag_runner.py": "031abf7d5b1a026eb199aa2a579bfb8baaa01175279d0018a364bfe50b76c3c1",
|
|
"tools/rag/evaluation/rag_metrics.py": "3e95922df90dd1b757a550040a2d9bd4f857e17d179f60170d7cc881a08f94b6",
|
|
"tools/answer_evaluation/context_policy.py": "e7b7c511e5e9cafdef04969eb4007a6a61ac4b99bcc60e9e722945a15bec1f26"
|
|
}
|
|
}
|