jaredpalmer/Kev 4B

• License: Apache-2.0
• Release date: 2026-09-24 (Kev 1.0 checkpoint, as documented in the original model card)
• Modalities: Input: text; Output: typed decisions (choice, score, noul)
• Original model: jaredpalmer/kev-4b

Description

Kev reads a document (the state) and typed questions, returning calibrated probabilities over the supplied options in one forward pass without generating text. This model uses a LoRA adapter and a pointer head on Qwen3.5-4B-Base. It is intended for document classification, routing, triage, and checks where probability thresholds can direct uncertain cases to human review.

Run

self serve kev:4b

Here is a link to the API docs: Decision API.

Example Request

With the server running, open another terminal and send a request:

curl http://localhost:8080/v1/systemone \
	-H 'Content-Type: application/json' \
	-d '{
		"state": "I was charged twice for the same order.",
		"questions": {
			"department": {
				"type": "choice",
				"instructions": "Which department should handle this support ticket?",
				"criteria": {
					"billing": "Payments and refunds",
					"technical": "Technical problems"
				}
			}
		}
	}'

The server returns a JSON decision for the department question.

Engine Parameters

Served by the upstream laya daemon through the ggmlc-laya adapter.

Parameter	Default	What it controls
threads	4	Number of CPU workers; accepts integers from 1 to 1024.
cuda_graph	true for auto / CUDA devices; otherwise false	Capture a CUDA graph for the live request shape.

Override startup settings with --set <parameter>=<value>.

Quantizations and Files

Default variant: q4. Sizes and SHA-256 hashes are pinned in the model registry.

q4 (default)

File	Role	Size	SHA-256	Origin
kev_4b_ud_q4_k_m.gguf	model	3.9 GiB (4211035872 bytes)	f2b31aa5a3c0f75f22f034f0452a8d7553d583c77b2783926f3ba1682143b92b	mys/kev-4b-GGUF

Total download: 3.9 GiB.

q8

File	Role	Size	SHA-256	Origin
kev_4b_q8_0.gguf	model	4.2 GiB (4498493152 bytes)	51c9353a5ff0b8e2f7cd7262359c0b26ea54b327214fe26a5da52b505b2978e8	mys/kev-4b-GGUF

Total download: 4.2 GiB.

fp16

File	Role	Size	SHA-256	Origin
kev_4b_f16.gguf	model	7.9 GiB (8441712352 bytes)	72fa1c764c99b9683d82333943224fceba094a3bce2426a651f2e23c34a9127e	mys/kev-4b-GGUF

Total download: 7.9 GiB.

Notes

• Text only; up to 16 options per question.
• Context and option limits are compiled into the GGUF and cannot be changed at runtime.
• Download size is not the total runtime memory requirement.
• Origin links use main because these registry entries pin hashes, not repository revisions.

Quality benchmarks

Decision Index: #84 (39.50)

jaredpalmer/kev-4b

Published result for the named upstream model and evaluated configuration; checkpoint revision and local quantization equivalence are not guaranteed.

Evaluated: kev-4b-r10

Edition: Decision Index 0.3.2

Checked 2026-10-13

Typed Decisions: Publisher-reported

jaredpalmer/kev-4b

Winnow publisher comparison of Kev-4B BF16 using pinned author runtime, 400 cases / 2000 decisions. Evaluated release may differ from current upstream; teacher agreement, not correctness or an official rank.

Dataset	Metric	Score
LocalLLaMA/typed-decisions (test)	Teacher agreement (%)	65.75

Hugging Face model-card evaluations: Publisher-reported

jaredpalmer/kev-4b

Publisher-reported upstream checkpoint; not an independent rank or local quantization

Checked 2026-10-13

Dataset	Metric	Score
transfer-v4 test: six never-trained public sources and held-out policy structures (656 questions)	accuracy	0.84
transfer-v4 test: six never-trained public sources and held-out policy structures (656 questions)	brier_score	0.22
breadth-v1 test: 14 held-out public datasets (3,089 questions)	accuracy	0.69
tasksource-heldout-v1 development: 17 of 24 held-out task families (1,993 questions)	accuracy	0.68
hard-v1 test (1,088 questions; programmatic labels, held-out templates)	accuracy	0.80
hard-v1 test (1,088 questions; programmatic labels, held-out templates)	brier_score	0.28
devtools-v1 test (1,071 questions; six public developer-tooling sources)	accuracy	0.76
devtools-v1 test (1,071 questions; six public developer-tooling sources)	brier_score	0.34

Benchmarks

No submitted benchmark reports yet.