Senova-4B-v4
Senova-4B-v4 is a full-parameter option-set soft-label fine-tune of
Qwen/Qwen3.5-4B on 5,181,618 examples for 3 epochs. This release is the
final checkpoint, step 1530. Model weights are licensed under Apache-2.0;
see LICENSE and NOTICE. Dataset rights are not granted by the model license.
Public Decision Index 0.3 index: 67.02, complete=true;
140,178/140,178 scoreable requests successful. The scored run uses Read Twice
and thinking off.
Inference repeats the complete rendered question input twice, separated by a
blank line, inside one user message, then reads one next-token distribution.
It does not make two model calls. Thinking stays off. The same Read Twice
setting is required for private evaluation. The Senova-4B-v3 inference server
can load these weights.
The weights have been modified by soft-label training and are not an unmodified
Qwen release. This score is a public benchmark result. The private Full score
is determined by the leaderboard maintainers.