jeba-multi (System One decision engine)
Status: released (
v0.2.0). Trained on a single RTX 3060 12GB and published as LoRA adapters (munod/jeba-en,munod/jeba-multi); measured numbers below come frombenchmarks/report.md.
Model details
- Developed by: The jeba Authors.
- Model type: non-autoregressive encoder with three task distributions (
noul,choice,score), answering typed questions about a state in one forward pass. - Trunk: ModernBERT-large (English) and mmBERT-base (100+ languages); see ADR-0007.
- Adapters:
munod/jeba-en,munod/jeba-multi(LoRA; load base + adapter). - Licence: Apache-2.0.
- Repository: https://github.com/munod/jeba
Uses
jeba answers atomic choice / score / noul questions about a state and returns typed values
with probabilities and confidence. It speaks the TypeSafe Jev /v1/systemone wire protocol as
a drop-in and runs locally/offline with no API key. Compose several atomic answers in code
rather than asking one broad question.
Out of scope: free-form text generation, multi-step reasoning, and any decision requiring extended deliberation — decompose those into atomic questions and combine results in code.
Bias, risks, and limitations
- Probabilities are only meaningful after calibration; the shipped temperature must be
applied (see
docs/training.md). - Synthetic training data can inherit generator biases; public probes are evaluation-only.
- Confidence is a property of the distribution, not a guarantee of correctness.
Training
Deterministic synthetic JSONL (training/generate_data.py) supervised with an RLCD
proper-scoring objective (training/finetune_rlcd.py), then temperature-calibrated on a held-out
split (training/fit_calibration.py). Configs and seed live under training/configs/.
Evaluation
Reported by training/evaluate.py and rendered by benchmarks/report.py (accuracy, ECE, p50/p95
latency per primitive and language).
Full-scale run (single RTX 3060 12GB): 9,000 English / 18,000 multilingual train / 1,500 eval
deterministic synthetic records (fully localized per language, a learnable other team with rich
descriptions, per-record RNG, one-in-six distractor clauses), LoRA (r=16) plus a dedicated
low-rank choice head (r=32, near-identity init), 4 epochs, batch 16, bf16 + gradient
checkpointing.
| Checkpoint | Accuracy | ECE (calibrated) | p50 (ms) |
|---|---|---|---|
| English (ModernBERT-large + LoRA r=16 + choice head) | 0.763 | 0.061 | 22.6 |
| Multilingual (mmBERT-base + LoRA r=64 + choice head) | 0.853 | 0.038 | 13.3 |
Per primitive (English): choice 0.834, noul 0.744, score 0.712; (multilingual): choice
0.684, noul 0.960, score 0.916. The localized, per-record-RNG data (B-1) lifted multilingual
choice from 0.40 to 0.68 and English overall from 0.72 to 0.76. Raising the multilingual LoRA
rank from 16 to 64 (alpha 128) removed the cross-language capacity bottleneck: overall accuracy
0.702 → 0.853 and es ECE 0.170 → 0.038 (es accuracy 0.472 → 0.956). Five of six languages now
meet ECE ≤ 0.05; nl (ECE 0.104, accuracy 0.663) remains the outlier (NFR-C06 partially open).
The CUDA-graph fast path (JEBA_FAST=1) gives a 2.7× p50 speedup with 0 top-label flips.
Robustness (B-4). On a noisy view (one surface edit — typo/accents/casing — applied to 15% of states) English drops only 0.763 → 0.760 and multilingual (r=64) 0.853 → 0.847, so the released adapters are already robust to this noise model.
Full tables and environment are in
benchmarks/report.md.
Citation
@misc{jeba2026,
title = {jeba: a local-first System One decision engine},
author = {The jeba Authors},
year = {2026},
howpublished = {\url{https://github.com/munod/jeba}}
}