Decision4Rec releases Kev models that rerank recommender candidates
Apache-2.0 adapters from 0.8B to 27B that read a pool of 30 candidates and output a probability for each. Development results only.
Marco Polignano and co-authors published decision4rec-kev-9b on Hugging Face on 19 October 2026, by repository creation date. Sibling repositories at 0.8B, 4B and 27B appear on the Hub. The card describes a decision model for recommender systems. It reads the 30 best candidates from a multi-agent pool, with item text, knowledge-graph statements and the agents that found each candidate, then outputs one probability per candidate in a single forward pass.
It uses rank-16 LoRA adapters plus a pointer head on a frozen Qwen/Qwen3.5-9B-Base, starting from the jaredpalmer/kev-9b adapter. The adapter and head are Apache-2.0; the base keeps its own licence. Training was one epoch on 8,000 sampled records from Amazon Video Games, Amazon Musical Instruments and Steam.
On development users the card reports NDCG@10 of 0.0736 on Video Games, 0.0485 on Musical Instruments, 0.0798 on Steam and 0.0406 on Industrial and Scientific, which was unseen in training. The pool-plus-ranker setting without the decision model scores 0.0730, 0.0468, 0.0786 and 0.0386.
The author notes these are development results only, with a confirmatory test planned. No inference provider hosts it.
A decision model that ranks, not routes. The margins are small and the card says so.
Sources
Editor's first pass (AI-assisted)
- Real 3.5: Adapters are public; sizes 0.8B, 4B, 9B and 27B exist on the Hub.
- Useful 3: Gains over the baseline are small, about 0.0006 to 0.002 NDCG@10 over the no-decision-model setting.
- Notable 2.5: Narrow research release.
- Novel 4: Applies the decision format to recommender candidate ranking.
Comments