kev-4b

kev-4b — research preview

kev-4b is a decision model: one document (the state) and a set of typed questions in, a probability distribution per question out, in one forward pass. No text generation. It is a LoRA adapter (r=16) plus a pointer head on Qwen/Qwen3-4B-Base, serving TypeSafe's public /v1/systemone contract.

Research preview, not a versioned release. It is the best 4B checkpoint under a frozen, checksummed protocol after ~40 controlled 4B trials, and the first kev whose out-of-domain accuracy is within ten points of Jev on the same items. This checkpoint clears the release screen we set in advance (held-out policy pairs 0.73 both-correct, screen 70%), but the other two seeds of the same recipe do not (0.62, 0.67), and our rule is every seed; so it stays a preview.

• Hub: jaredpalmer/kev-4b (this repo; trial v7-rc3/01-trial-1)
• Code, suites, every trial with hashes and paired bootstraps: github.com/jaredpalmer/kev — PLAN.md, runs/leaderboard.md

Results (same frozen items for every row)

Per-source out-of-domain accuracy (kev-4b / Jev): QNLI 0.89 / 0.93, SciQ 0.99 / 0.99, TweetEval-offensive 0.75 / 0.81, PAWS 0.72 / 0.79, MMLU 0.65 / 0.90, Emotion 0.66 / 0.59, deadline (3-level date arithmetic) 0.53 / 0.93, (A and B) or not C 0.97 / 0.97, if A then not B else C 0.88 / 0.78.

Seeds: three seeds on decision-v7: transfer 0.773 / 0.790 / 0.770, held-out pairs 0.62 / 0.73 / 0.67; this checkpoint is seed 1, selected on development transfer accuracy. Trained on decision-v7 (10k public records + 896 policy records over nine template families incl. four ordinal Score threshold families + 1,680 records from 60 random rule structures with negation anywhere); development/test items are byte-identical to v4, so every number here is comparable with earlier previews.

Locked test, one exploratory read (runs/locked/kev-4b-v7-preview-ungated/, labelled ungated because the screen is not met on every seed): in-distribution 0.856 (Brier 0.211), out-of-domain 0.806 (Brier 0.294, confident errors 6.6%, held-out pairs 0.66). This partition will not be read again for this checkpoint.

What we learned building it

• Capacity dominates out of domain. With public examples and synthetic budget held equal, 0.6B → 4B is +14–19 pp; 4B → 8B is +1–7 pp.
• Fine-tuning erodes base capability, and the learning rate controls it. The 4B base, zero-shot with a letter readout, scores 0.688 on the same MMLU items and 0.787 on PAWS; the default recipe (lr 2e-4) trained down to 0.60–0.66 / 0.56–0.71. Lowering lr to 5e-5 recovers most of it and is the single largest recipe improvement we found; fewer LoRA target modules and smaller ranks help less.
• More public training data raises in-distribution accuracy and lowers transfer at 4B (10k vs 3.4k records: −3 pp). Knowledge MCQ sources (ARC, OpenBookQA, CommonsenseQA) raise in-distribution accuracy to 0.86 without moving transfer.
• Programmatic contrastive policy pairs teach the trained rule structures (both-correct 0.85–1.0) but transfer to unseen structures only partially (0.5–0.6 at 4B, 0.03–0.11 at 0.6B).

Known limits

• Held-out policy reasoning (unseen rule compositions, date arithmetic with grace periods) is far from Jev.
• Product-shaped questions with no training analogue are not guaranteed: on the TypeSafe docs example ("two charges on my card" → Is there a billing problem?) this checkpoint answers 0.22 while kev-0.6b answers 0.97 and picks the return reason (wrong size, 0.54) correctly. Lower drift from the base means fewer task-specific priors; measure on your own inputs.
• Out-of-domain probabilities are usable but not calibrated (raw ECE 0.096); temperature fitted in-domain does not transfer.
• 4B fp32 needs ~16 GB; on a 32 GB Mac use KEV_DTYPE=bf16. Latency on an H100 is ~45 ms per packed request; on an M5 several hundred ms.

Training

Frozen suite evals/v4/decision-v4: 10,000 public records (1,000 per source, ten sources) plus two programmatic policy arms of 448 records, two epochs, LoRA r=16 on attention and MLP projections, pointer head from scratch, cross-entropy on the option distribution, lr 5e-5 (OneCycle), effective batch 8, bf16 autocast with fp32 master weights, gradient checkpointing, one H100 (~40 min). Augmentation: option permutation, none-of-the-above insertion, distractors, none minimal pairs on 25% of Choice records. No Jev outputs were used for training.

Evaluation protocol

Development partitions select models; the locked test partition is read at most once per candidate. Every number carries suite hash, code hashes, and git commit in result.json. See PLAN.md for the corrections we made to our own earlier claims.

Use

uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8008      # KEV_DTYPE=bf16 on a 32 GB Mac

Any TypeSafe-compatible client works: TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8008", model="kev-latest").

License

Apache-2.0 for the adapter and head; Qwen3 base is Apache-2.0; datasets carry their own licenses.