Tasksource ships JEV-Nano-v0, a 149M multi-vector typed-decision model

Source · Model card: Tasksource-JEV-Nano-v0

On 28 September 2026, Damien Sileo published Tasksource-JEV-Nano-v0 on Hugging Face: a ~149M fine-tune of LateOn that answers typed choice, noul and score questions with ColBERT-style MaxSim scoring instead of a classifier head.

The situation (state + question) is encoded once and reused across candidate sets; options are scored independently, and the card claims exact permutation equivariance. Training used the companion tasksource-jev-typed-decisions stream (512k decisions, 519 tasks, no teacher). Inference is via pylate + torch — there is no /v1/systemone server. Weights Apache-2.0; the dataset carries HF license tag other.

What the author reports (UnverifiedClaim): Typed Decisions 42.6%, Banking77 44.7%, classifier-benchmark v2 57.5 micro, and a peer table against GLiClass / GLiNER2.5 on the same 1k samples. Learned temperature 0.53. No public p50/p99.

Catalog entry: /models/tasksource-jev. Nearby peers: Julia-1, GLiNER2.5-Decide, Laya.