673 3 weeks ago

Typed decision model. Recommended file: Kev-4B-Artemis-Q4_K_M.gguf. Use /v1/systemone. Measured Q4_K_M: same answer as BF16 93.0%, 0 confident answers changed on typed-decisions test split (selection panel); protocol kev4b-rtx5090-20261012.

ollama run turintech/kev-4b-artemis

Models

View all →

Readme

Kev-4B Artemis

Kev-4B for routing, triage and quick checks on your own GPU. It answers typed questions (yes or no, pick one, a score) with a probability for each answer, in one pass.

  • 2.54 GB download for the default file, Kev-4B-Artemis-Q4_K_M.gguf.
  • Same answer as the original on 93.0% of 2,000 test decisions, against 93.2% for ggml-org’s Q4_K_M (3.03 GB); 0 confident answers changed. These decisions also chose the file.
echo "My Stripe integration keeps failing. Please help ASAP." | ollama run turintech/kev-4b-artemis:2.5gb --ask "Does this message express urgency?"

Kev-4B is a 4B decision model by Jared Palmer.

From code

Ollama serves /v1/systemone with a model field added.

curl http://localhost:11434/v1/systemone -d '{
  "model": "turintech/kev-4b-artemis:2.5gb",
  "state": "My Stripe integration keeps failing. Please help ASAP.",
  "questions": {"urgency": {"type": "noul", "instructions": "Does this message express urgency?"}}
}'

Settings

The package sets num_ctx 16384 and num_batch 4096, the card’s llama-server settings (-c 16384 -ub 4096 -b 4096). The batch size moves the probabilities slightly, so keep it when you compare with the card.

Results

Our default file against the closest public file and the original, on 2,000 typed decisions from the test split. These decisions were also used to choose the file.

File Size Same answer as original Confident answers changed Accuracy
Ours, Q4_K_M 2.54 GB 93.0% 0 66.8%
ggml-org Q4_K_M 3.03 GB 93.2% 0 66.8%
BF16 original 8.43 GB reference reference 66.8%

The public Q4_K_M is larger but stays closer to the original on the 1,000 train-split decisions (94.0% against 92.7%). How it was tested.

Files

Every file we publish for this model. Sizes are download sizes, not memory.

File Size Use it for Ollama tag
Kev-4B-Artemis-Q4_K_M.gguf 2.54 GB Default q4_k_m (latest)
Kev-4B-Artemis-IQ3_M.gguf 2.17 GB Smaller: a little less memory iq3_m
Kev-4B-Artemis-Q3_K_M.gguf 2.04 GB Smaller: tight memory; about 2 to 3 points lower accuracy q3_k_m
  • The Ollama tag runs our package, which holds the same file: ollama run turintech/kev-4b-artemis:<tag>.

Limitations

  • The additional panel historically called held-out has not been verified independent of calibration.
  • Small differences do not establish general improvements.
  • The numbers come from llama.cpp, not from Ollama or LM Studio.

Licence

Apache-2.0, the same as jaredpalmer/kev-4b. The package carries the licence and TurinTech’s NOTICE; the Hugging Face repository has both files.