StackMap
Subscribe
Explore / kev
jaredpalmer

kev

Open reproduction of TypeSafe's Jev: a LoRA + readout head on Qwen that answers many typed questions about one document in a single prefill pass, returning calibrated probabilities.

492 30 Python Apache-2.0updated today
traininglocal
View on GitHubDispute this mapping →
Curator's take

Reach for kev when the task is decisions, not prose: route this ticket, score this risk, answer 30 yes/no questions about one document — and you want a number you can threshold on rather than text you have to parse. The head is trained with cross-entropy on labelled outcomes, so the probabilities mean something (kev-8b: Brier 0.34, 8% confident errors out of domain), and a block-causal mask keeps questions from seeing each other, verified to 4e-6 against separate requests. It speaks TypeSafe's System One API, so their SDK works against localhost with a base_url change. Not a general chat or agent model - it never decodes, and it only answers the question types you define (noul/choice/score). Don't use it zero-shot on your own domain either: the value is in training the head on your labels, and the three 0.6B/4B/8B checkpoints are explicitly previews that failed the author's own release screen on held-out rule reasoning. Real Jev still wins out of domain (0.86 vs 0.77) - kev's pitch is that it runs on your laptop and you own the weights.

Mapped by ShipWithAI editors · links verified

Continue your stack

What teams reach for next — and why each earns a place beside kev. Ranked by curator confidence.

pairs wellpairs wellpairs wellalternativealternativealternativealternativealternativeAdalajev-ultrafastfast-jev-compactionlocaljevcontextgemneedleDSPyMLX-LoRA-Studiokev
pairs wellalternativebuilt withpick a node for the why · open it from the panel
Weekly digest
README.md2 min read

kev

Jev-inspired decision model. Typed questions in, calibrated probabilities out, one forward pass.

CI Weights: kev-0.5b · 0.6b · 4b · 8b Frozen eval suites Research log License: Apache-2.0

kev playground

kev is a LoRA adapter and a small readout head on top of a Qwen base model (0.5B to 8B). It reads a document once and answers many typed questions about it in parallel, in a single prefill pass with no decoding. The document and every question are packed into one sequence; a block-causal mask lets each question see the document but never another question. A pointer head then scores each question's options against its decision token and applies softmax. Those probabilities are the output. The head is trained with cross-entropy against labelled outcomes, so the probabilities are learned rather than generated as text.

The architecture follows the reconstruction of TypeSafe's Jev in Jev's Architecture Unmasked. The API follows TypeSafe's System One contract, so the official typesafe-sdk works against a local kev server with a base_url change.

Highlights

  • Three question types. noul (yes/no), choice (2–255 options), score (ordered levels). One shared readout.
  • One pass, many answers. The state is encoded once. Questions run as isolated branches under a block-causal mask.
  • Isolation is exact. A question cannot see a sibling question. Packed and separate requests agree to 4e-6.
  • Probabilities, not prose. Trained with cross-entropy on labelled outcomes. Out of domain, kev-8b has Brier 0.34 and 8% confident errors on sources it never saw.
  • Drop-in API. POST /v1/systemone with TypeSafe's request and response shapes. Their SDK's quickstart runs unmodified.
  • A family, measured the same way. 0.5B, 0.6B, 4B and 8B checkpoints scored on frozen, checksummed suites with a locked test, against the real Jev on the same items. Out of domain: kev-4b 0.76, kev-8b 0.77, Jev 0.86.
  • Runs on a laptop; trains in the cloud. kev-0.5b trains in ~1h45m on an Apple M5; the 4B/8B recipes train in 40–70 min on one H100 via Modal and serve on a 32 GB Mac in bf16.

kev family vs Jev on sources kev never trained on

Installation

Requires Python 3.12+, uv, and Node 20+ for the playground. Serving is tested on Apple Silicon (MPS); training and evaluation on CUDA (H100 via Modal) and MPS.

git clone https://github.com/jaredpalmer/kev.git && cd kev
uv sync --extra serve
cd playground && npm install && cd ..

Download the weights

All checkpoints are on the Hugging Face Hub in the kev collection. --run accepts a Hub id; the base model downloads on first load. kev-4b is the one to start with: the best accuracy per byte, and it serves on a 32 GB Mac in bf16.