use cases
Is it possible to use it as a memory-reranker?
@repla73 A decision model can be used as a pointwise memory reranker: put the query and one retrieved memory in the state, ask how relevant the memory is, and define a 0–4 relevance scale. You can rank memories by the expected score from the option probabilities, rather than only the winning label.
I built another local decision model, Seb-9B, with this kind of scoring interface: https://huggingface.co/ironbcc/seb-9b. It runs through vLLM or MLX, with GGUF builds available. I haven't evaluated it as a memory reranker yet, so this is a proposed use case, not a benchmark claim. I'd retrieve a small candidate set with embeddings first, then compare reranking against a dedicated reranker using nDCG@k and end-to-end latency. Each query/memory pair requires its own score.