CanadaOrNaw/dex-1b

Dex

Dex: an open-source, Jev-compatible, non-autoregressive System 1 decision model. Typed choice, score and yes/no decisions over any text in a single forward pass. It never generates text, so there is nothing to parse and nothing to hallucinate. Ettin-1B bidirectional encoder (1.1B parameters, 784M fine-tuned) with a decision head trained from scratch, trained against strictly proper scoring rules. Same POST /v1/systemone request and response shape as TypeSafe Jev. Apache 2.0.

Use

pip install laya torch transformers
huggingface-cli download CanadaOrNaw/dex-1b --local-dir dex-1b

import torch
from laya.agent import Agent
agent = Agent("dex-1b", device="cpu")
agent.model.encoder.to(torch.bfloat16)
agent.model.to("cuda"); agent.device = torch.device("cuda"); agent.dtype = torch.bfloat16
agent.system_one("Hi, I was charged twice. Please refund the duplicate.",
    {"refund": {"type": "noul", "instructions": "Is the customer asking for money back?"}})["answers"]

A ready HTTP server with the Jev request/response shape (POST /v1/systemone), a systemd unit, a Dockerfile and tests are in the GitHub repository: https://github.com/CanadaOrNaw/dex

Training

Trained against strictly proper scoring rules (cross-entropy on the chosen option plus KL divergence to the target probabilities). Ettin-1B encoder, 784M of 1.1B parameters fine-tuned (embeddings and the lowest 7 of 28 layers frozen), decision head trained from scratch; 3 epochs, bfloat16, batch 64, encoder learning rate 4e-5, head 2e-4, 30% option-order shuffling. Public text and instruction-task collections with about 1,080 distinct task instructions; see DATA_CARD.md.

Results (accuracy against the human label)

accuracy against the human label	questions	JEV	Dex	Stock Laya
A. Ten most-cited JEV use cases	1,244	84.6%	81.5%	68.2%
B. Tasks never trained on	1,737	84.5%	74.4%	63.0%
C. Exam: 8 datasets never used	2,346	78.2%	63.8%	62.5%
All three	5,327	81.8%	71.4%	64.0%

Speed on an RTX 3070 Ti: 18 ms for one question on a short text, 95 ms for fifteen; 2.4 GB GPU memory in bfloat16. Picks a different option when the answer order is reversed 7.0% of the time (stock Laya 25.2%, JEV 3.0%). Per-dataset tables, calibration, ablations and failure analysis: BENCHMARKS.md in the GitHub repository.

Jev Decision Index (self-run)

Self-run through the edition-0.3 harness on 2026-10-09 (140,620 requests, RTX 3070 Ti): skill 12.99, raw 32.47 — about 3x stock Laya's 4.43 on the official board, mid-pack of the sub-2B class. Not an official board score (the board mixes in private tests only the maintainer can run); Dex is not on the board until it is evaluated there. Full per-area breakdown: BENCHMARKS.md.

Limits

• No knowledge it was never given. On medical multiple choice (MedMCQA) Dex scores 30.7% against JEV's 78.7%; on reading an abstract (PubMedQA) 55.0% against 78.3%. More training did not change this: the gap is knowledge, not format.
• Stock Laya is better at some screening tasks. Jailbreak screening: stock Laya 98.7%, Dex 89.3%. Moderation and spam: stock Laya 89.9%, Dex 83.0%.
• It cannot judge whether code is correct (52.5%, a coin flip; JEV 90.8%), and it is behind on answer verification (90.0% against JEV's 96.9%).
• Not a drop-in replacement for JEV in every workflow. In our own multi-step automation checks it was well behind JEV; use it where the benchmark above shows it is close.
• Slower than Laya for many questions or long texts, still far faster than the Jev API: 15 questions on a short text take 95 ms (Laya 47 ms); on a ~900-token text 878 ms (Laya 260 ms); the Jev API takes 780 to 890 ms whatever the size. It needs 2.4 GB of GPU memory in bfloat16.
• English only, trained with inputs up to 1,792 tokens and up to 30 options per choice question.
• Single training run. We did not measure run-to-run variation.
• Our benchmarks. We built them, scored every system on the same prompts, and report intervals, but benchmark A shares datasets with the training data (different rows) and its routing suite uses a proxy answer key. B and C do not.
• Probabilities are temperature-fitted, not guaranteed calibrated. Calibration error 0.058 across our tests (JEV 0.052, stock Laya 0.092); when Dex is 90%+ sure it is right 93.2% of the time, but it is that sure on only 29% of questions (JEV 63%).
• Decision Index coverage gaps. About 11% of index requests come back unsupported (long texts past 1,792 tokens; API-Bank entirely unanswered), and the index counts unanswered as wrong. That coverage gap, not accuracy, is the biggest drag on the index score above.

Licence

Apache 2.0. Built on the Laya runtime (Apache-2.0) and the Ettin-1B encoder (MIT). Data and benchmark credits are in the data card. Not affiliated with TypeSafe AI or Convai Innovations.