cbjev

Typed decisions — choice, score, noul — about any text or JSON state, from one encoder pass. Every question of a call shares one encoding of the state. Code, benchmarks and training pipeline: https://github.com/tomek7667/cbjev · project page: https://tomek7667.github.io/cbjev/

checkpoint	path	encoder	use it for
cbjev	repo root	ModernBERT-large (395M)	English
cbjev-multilingual	multilingual/	mmBERT-base (307M)	other languages (the Router picks it for you)

Both are fine-tuned from the Apache-2.0 Laya checkpoints (laya-typed-decisions and laya-multilingual) on 35 public datasets, in cbjev's shared layout, then calibrated on a held-back dev split. They need the cbjev runtime; transformers alone will not run them.

pip install "git+https://github.com/tomek7667/cbjev"

import cbjev
agent = cbjev.load()           # downloads this repo on first use
agent.predict({"body": "Billed twice for March. Refund today or we cancel."}, {
    "team": {"type": "choice", "instructions": "Which team should handle this?",
             "criteria": {"billing": "invoices, refunds", "technical": "bugs", "other": "anything else"}},
    "churn": {"type": "noul", "instructions": "Does the customer threaten to cancel their subscription?"},
})

Results (RTX 4090, same cases for cbjev and Laya; Jev = published third-party numbers)

	cbjev	Laya (best checkpoint)	Jev (published)
mean accuracy, 15 English suites	0.741	0.710	—
typed-decisions, 2,000 decisions	0.783	0.768	0.727
mean ECE (lower is better)	0.117	0.125	0.246
answer flips when options are reordered	0.2 %	7.8 %	13 %
MASSIVE intent, 51 languages	0.436	0.401	—
10 questions, 500-token document	11.4 ms	75.8 ms	—

Not better everywhere: cbjev trails Laya's better checkpoint on AG News (−0.8), DAIR emotion (−2.5), prompt injection (−3.4) and support triage (−4.0), and Jev's published Banking77 (0.870 vs 0.620). Full tables: BENCHMARKS.md.

License

GPL-3.0-or-later. Fine-tuned from Apache-2.0 Laya checkpoints by Convai Innovations — see NOTICE.