Open release: ZooWork Instinct Tuned 4B
#1
by pierrexsq - opened
We’re releasing ZooWork Instinct Tuned 4B, an Apache-2.0 decision model built on Qwen3.5-4B.
Instead of generating text, it scores candidate-label logits in one forward pass and returns probability distributions for choice, noul, and ordinal score questions. The release includes merged BF16 weights, the decision prompt and temperature configuration, file hashes, and an open reference runtime.
On the 231 published JevBench tasks it scores 198/231 (85.71%). These public tasks were used during development, so this is a transparent reference result—not a held-out claim.
- Model card and weights
- Open-source runtime
- Hosted API using model ID
instinct-tuned-4b
We’d especially welcome feedback on domain transfer, calibration, and option-order sensitivity.