juteq/arbit-base

Arbit base checkpoints

Starting weights for Arbit, a non-autoregressive decision engine: one forward pass answers several typed questions (choice, score, noul) about a shared state and returns a calibrated probability distribution for each.

path	checkpoint	encoder	context
repo root	English general	ModernBERT-large (421M)	512 tokens
typed-decisions/	fine-tuned on the typed-decisions benchmark's train split	ModernBERT-large (421M)	1024 tokens

import arbit
agent = arbit.load("juteq/arbit-base")                       # English
tuned = arbit.load("juteq/arbit-base", subfolder="typed-decisions")

Provenance

These files are redistributed unmodified from a publicly released Apache-2.0 checkpoint; JUTEQ did not train them. They are published here so that training runs that start from them are reproducible. The encoder is answerdotai/ModernBERT-large (Apache-2.0).

Models trained by JUTEQ are published separately.

Known limits

• English only.
• typed-decisions/ is fine-tuned on four synthetic workflows; its scores measure agreement with that benchmark's teacher model, not correctness.
• Probabilities need temperatures fitted on your own held-out data; the shipped ones were fitted elsewhere.

License: Apache-2.0 (see LICENSE in this repository).