Question Answering on HotpotQA distractor (val)
53.53Answer AccuracyProvenAI
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| ProvenAIRetrieval depth=10, Embedding model=all-MiniLM-L6-v2, Index=FAISS IP + SQLite, Generation backend=MLX (Apple Silicon), Generation model=Qwen2.5-3B-Instruct, Citation limit=2, Prompt doc limit=900 chars, Threshold τ=0.34, Influence weights=0.8 / 0.22026.06 | 53.53 | 71.55 | 71.55 | 71.55 | 0 |