Jared Palmer's Kev-0.5B Answers Many AI Questions in one 38MB Pass
A LoRA-tuned Qwen2.5-0.5B that reads a document once and answers many typed questions in parallel with calibrated probabilities, no text generation involved.
- Kev-0.5B is an open-source LoRA adapter on Qwen2.5-0.5B that reproduces TypeSafe's Jev decision-model architecture.
- Reads a document once, answers many typed yes/no, choice, and score questions in parallel in one prefill pass with no decoding.
- Achieves 0.799 accuracy and 0.065 ECE across six held-out datasets, dropping to 0.031 ECE after temperature scaling.
- Trains in about 1h45m on an Apple M5 laptop, weights are only 38 MB, released under Apache-2.0.
- Drop-in compatible with the official
typesafe-sdkvia abase_urlchange to a local Kev server. - Architecture based on Archer Hume's reverse-engineering; code at github.com/jaredpalmer/kev.
Kev-0.5B packs many typed decisions into one model pass
Jared Palmer has released Kev-0.5B, an implementation of the architecture inferred from TypeSafe’s proprietary Jev system. It reads a shared document once and returns probability distributions for many typed questions in a single forward pass, giving classification, routing, moderation, and scoring pipelines an alternative to autoregressive JSON generation.
When a chat-completion API emits a confidence such as 0.92, the token probabilities describe how likely the model was to produce that string; empirical correctness requires calibration against labeled outcomes. Kev trains its decision head with cross-entropy on those outcomes, allowing developers to measure and adjust the resulting distributions.
A 38 MB adapter with an API
Kev combines a LoRA adapter with a small pointer head on top of Qwen/Qwen2.5-0.5B. LoRA adds compact trainable matrices to a mostly frozen model, while the pointer head converts internal representations into scores over the supplied options. Palmer based the implementation on Archer Hume’s Jev architecture analysis.
| Component | Details |
|---|---|
| Backbone | Qwen/Qwen2.5-0.5B |
| LoRA parameters | 8.8 million |
| Pointer-head parameters | 0.46 million |
| Total trainable parameters | 9.3 million, or 1.9% of the backbone |
| Artifact size | 38 MB |
| HTTP interface | /v1/systemone |
| Type | Purpose | Returned value |
|---|---|---|
noul |
Yes-or-no decisions | P(yes) |
choice |
Selection among 2 to 255 options | Top option and probabilities |
score |
Ordered levels such as a 1-to-5 rating | Expected level |
A local Kev server implements Jev’s
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
