kev-0.5b

kev-0.5b — first checkpoint

LoRA adapter + pointer head on Qwen/Qwen2.5-0.5B for a Jev-inspired decision model. See MODEL_CARD.md for the full recipe, data, metrics and limitations.

Held-out (in-distribution), 1,350 questions: accuracy 0.799, ECE 0.065 (0.031 after temperature scaling T=1.47). Question isolation exact; packed vs separate max Δp 3.7e-6.

Contents of kev-0.5b.tar.gz

file	what
adapter_config.json, adapter_model.safetensors	peft LoRA adapter (r=16, α=32, all projections), 35 MB
head.pt	pointer head weights + {"base": "Qwen/Qwen2.5-0.5B", "lora": 16}
tokenizer.json, vocab.json, merges.txt, *_config.json, added_tokens.json, special_tokens_map.json	tokenizer as saved by the training run
eval.json	full evaluation output (same file as runs/kev/eval.json in the repo)
train.log	training log used for the README plot
README.md	copy of the model card

The base model is not included; it is downloaded from the Hub on first load.

Use

mkdir -p runs && curl -L https://github.com/jaredpalmer/kev/releases/download/v0.1.0/kev-0.5b.tar.gz | tar xz -C runs
mv runs/kev-0.5b runs/kev
uv run --extra serve python -m kev.serve --run runs/kev --port 8009

Verify: shasum -a 256 -c kev-0.5b.tar.gz.sha256

Training recipe

uv run python -m kev.train --n_per_source 1500 --epochs 2 --perm_kl 0 --ord_w 0 --out runs/kev

Plain cross-entropy with data-level option shuffling; the ordinal and permutation-KL terms now on by default in train.py were not used for this checkpoint. Trained 2026-09-17 on an Apple M5 (32 GB), ~1h45m, code at commit d0e2b1f–ac67bf4.

License: Apache-2.0 (adapter and head). Base model under the Qwen license.