Opal Recall
The memory search model in the IMPERIAL app: it finds the right saved memory, earlier message or book passage for what the user says, on the iPhone. BAAI/bge-small-en-v1.5 fine-tuned on 13644 examples. Use CLS pooling and normalise.
Training data: questions written by Qwen/Qwen3-14B-AWQ (Apache 2.0) about passages from public-domain Project Gutenberg novels, invented memories, and messages from practice conversations written by open models. No real user data.
Held-out scores (recall@1 / recall@5): {"memory": [0.348, 0.636], "book": [0.419, 0.694]}. Scores below 0.81 mean "not relevant".
- Downloads last month
- 24