rag
healthcare
omop
ehr
data-engineering
ehr-omop-rag-assistant / evaluation /final_results.json
mieraci22's picture
Prepare EHR to OMOP RAG Space
c221896
Raw
History Blame Contribute Delete
1.51 kB
{
"project_test_rows": 10,
"systems": {
"final_constrained_rag": {
"reader": "google/flan-t5-large",
"reader_revision": "0613663d0d48ea86ba8cb3d7a44f0f65dc596a2a",
"retriever": "sentence-transformers/all-MiniLM-L6-v2",
"retriever_revision": "1110a243fdf4706b3f48f1d95db1a4f5529b4d41",
"omop_target_accuracy": 1.0,
"ragbench_hit_at_1": 0.7375510204081632,
"ragtruth_hallucination_f1": 0.632,
"mirage_hit_at_1": 0.6779100529100529
},
"base_closed_book_flan_t5_large": {
"omop_target_accuracy": 0.5,
"ragbench_hit_at_1": 0.0,
"ragtruth_hallucination_f1": 0.518,
"mirage_hit_at_1": 0.0
},
"rag_flan_t5_small": {
"omop_target_accuracy": 0.3,
"ragbench_hit_at_1": 0.7375510204081632,
"ragtruth_hallucination_f1": 0.632,
"mirage_hit_at_1": 0.6779100529100529
},
"rag_flan_t5_base": {
"omop_target_accuracy": 0.6,
"ragbench_hit_at_1": 0.7375510204081632,
"ragtruth_hallucination_f1": 0.632,
"mirage_hit_at_1": 0.6779100529100529
}
},
"secondary_retrieval_metrics": {
"ragbench_recall_at_3": 0.7549523809523809,
"mirage_recall_at_3": 0.9129872134038801
},
"notes": [
"All RAG reader variants share the same MiniLM retriever and groundedness policy, so their public component-benchmark values are identical.",
"The final pipeline locks the structured target to the highest-ranked corpus metadata; FLAN-T5 writes the rationale only."
]
}