Jev directory
CommunityJev Models & Reimplementations
Laya FullFT Typed Decisions
jjshin's full fine-tune of Laya, the ModernBERT-large System-1 decision model, with a 421M encoder and a decision head trained by a hand-written 404-line PyTorch RLCD loop and LBFGS temperature calibration.
- Reading time
- 2 min
- Published by
- Community
- Author
- jjshin
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Full fine-tune of Laya: ModernBERT-large 421M encoder plus a 2-layer decision head, trained on LocalLLaMA/typed-decisions.
- Self-reported 8-suite A/B: anti-jailbreak accuracy 0.84 versus base 0.795 and overall ECE 0.098 versus 0.188.
- Emotion-suite ECE fell to 0.107 from 0.284; fitted temperatures are 1.101, 1.039 and 1.144 per decision type.
- Pure-PyTorch RLCD loop, 404 lines, no trainer framework: Gaussian logit noise, 4 copies per state, sigma 0.4 to 0.1.
- Trained 4 epochs and 7,313 updates in 1.96 h on a single RTX 4090 in bf16; base is convaiinnovations/laya.
Quickstart
python
sigma = 0.4 + (0.1 - 0.4) * epoch / (epochs - 1)
eps = torch.randn((4,) + logits.shape) * sigma * maskWatch out
Card reviewed 2026-10-17; the 0.84 anti-jailbreak accuracy, halved ECEs and A/B figures are self-reported from the card's internal 8-suite run (laya-eval-ab-1001) with no external reproduction. The card itself notes the proper-scoring loss is non-monotonic by design (0.297 to 0.693), the training is a single 4090 run over 4 epochs, and the base article's own eval is 0.362.
Questions about this entry
Short answers drawn from the linked project or page.
A full fine-tune of Laya, the System-1 decision model: a ModernBERT-large encoder plus a 2-layer decision head, trained on the 1,200 train items of LocalLLaMA/typed-decisions for millisecond-scale action selection.
Its decision head outputs a probability distribution over N action options via masked softmax, so it answers with calibrated probabilities rather than generated text, like Jev's typed decisions; the card cites the base article's evaluation at 0.362.
Four steps: the policy outputs a masked-softmax distribution, Gaussian noise is added to logits at four copies per state with sigma annealed 0.4 to 0.1, a proper reward combines log score, spherical score and RPS, and REINFORCE with a group-mean baseline updates the weights.
The training loop is hand-written PyTorch, about 404 lines: AdamW with grouped learning rates (encoder 2.5e-5, head 1e-4) and CosineAnnealingLR. The card argues swift and TRL have no trainer for calibrating an option distribution, so nothing could be reused.
Self-reported on an internal 8-suite A/B (run laya-eval-ab-1001): anti-jailbreak accuracy 0.84 versus 0.795 base, ECE 0.098 versus 0.188, and emotion-suite ECE 0.107 versus 0.284, with fitted temperatures 1.101, 1.039 and 1.144.
LocalLLaMA/typed-decisions (all/train, 1,200 decision items) over 4 epochs and 7,313 updates in 1.96 hours on a single RTX 4090 in bf16; the base is convaiinnovations/laya and the licence is Apache-2.0.
Exploration is external Gaussian noise on logits rather than self-sampling from a token distribution, and the reward is a proper scoring rule whose expectation is maximised at the true distribution, so the model cannot game confidence; a post-training LBFGS step fits one temperature per decision type.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
A LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
Jev Models & Reimplementations#community#open-models#local
Community
A bilingual Turkish and English 4B Jev-class decision model from Berkay Hasret: a roughly 130 MB LoRA adapter on Qwen/Qwen3.5-4B that answers up to 16 typed questions per text in one pass and reports 0.774 JevBench macro accuracy.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.