Kev-4B

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, 33.8M trainable parameters) plus a pointer head on Qwen/Qwen3.5-4B-Base (revision 1001bb4d), serving TypeSafe's public /v1/systemone contract.

This version (2026-09-24): real-document delta. The previous Kev-4B plus one epoch (lr 2e-5) on documents-v1 train: 5,219 real US consumer-finance complaint narratives (CFPB, 2015-2024, up to ~7k tokens) with 7,488 questions (which product, which main issue), labels kept only where two open-weight teachers agreed with the consumer's own filing, mixed with 2,000 replayed decision-v7 records. On complaint narratives it has never seen, accuracy goes from 0.804 to 0.904 on the locked test (+9.9 pp [+7.5, +12.4], 936 questions) and from 0.811 to 0.891 on a private held-out set (documents-v2, 953 questions); on the development split it scores 0.895 against Jev's 0.868. Everything else is unchanged within noise: locked out-of-domain test 0.835 (previous 0.837), served Brier 0.233 (0.232).

Read this before relying on the documents numbers. The gain is measured in distribution: training and every documents suite share one source (CFPB complaints) and the same two question templates. It shows Kev-4B learns real long documents from a few thousand labelled examples; it does not show the same gain on other kinds of documents. Evaluation labels are AI-adjudicated (a unanimous three-model judge panel, or two agreeing adjudications) and human spot-checked (47/50 and 50/50).

• Hub: jaredpalmer/kev-4b (this repo; trial r8-small/00-trial-0; the registration and every read are in PLAN.md round 8 on the research/overnight-r6 branch). The previous version is at tag night2-du-release; the pre-delta v7 checkpoint at v7-base; the Qwen3 generation at qwen3 (its card).
• Code, suites, every trial with hashes and paired bootstraps: github.com/jaredpalmer/kev. The numbers below are in runs/release/kev-4b-r8.json.

Results (as served: each checkpoint at its own fitted temperature)

	Kev-4B (this version, T = 2.96)	previous Kev-4B (T = 2.14)	Jev
real documents, locked test (documents-v1, 936 questions)	0.904	0.804	–
real documents, private held-out (documents-v2, 953)	0.891	0.811	–
real documents, development (920)	0.895	0.811	0.868
real documents, Brier (locked test)	0.156	0.286	–
in-distribution accuracy (decision-v7 dev, 1,264 questions)	0.873	0.872	0.845
out-of-domain accuracy (transfer-v4 dev)	0.802	0.797	0.857
out-of-domain Brier / ECE	0.265 / 0.043	0.264 / 0.040	0.211 / 0.049
confident errors out of domain (p ≥ 0.9 and wrong)	2.9%	2.6%	3.7%
coverage at ≤ 5% error	0.552	0.573	0.70
held-out policy structures, both siblings correct	0.781	0.781	0.86
unknowable items answered at ≥ 0.9 (transfer-v9)	0.00	0.00	0.09
MMLU-Pro (transfer-v9 dev, 10-way)	0.515	0.490	0.840
locked test, out-of-domain accuracy / Brier	0.835 / 0.233	0.837 / 0.232	–
locked test, in-distribution accuracy	0.875	0.871	–
SemIf (144 authored decisions)	0.882	0.889	–
scienthoon (873 support tickets)	0.723	0.696	–
WANLI-v2 (1,002 NLI pairs)	0.691	0.699	–
TypeSafe (89 answered rows)	0.843	0.843	–

Paired against the previous version (record-clustered bootstrap, 95 %): documents dev +8.4 pp [+6.0, +10.6], locked test +9.9 [+7.5, +12.4], private held-out +8.0 [+5.6, +10.3]; scienthoon +2.6 [+0.8, +4.4]; SemIf −0.7 [−2.8, +1.4]; WANLI-v2 −0.8 [−2.0, +0.4]; TypeSafe identical on all 89 rows. The release was decided by a rule registered before any read (PLAN.md round 8, research/overnight-r6 branch): documents development lower bound > 0, short-state and external guards sized to what each suite can resolve, then one read of the untouched documents test and one locked read. A first seed (round 7) gave the same documents gain and failed only per-suite lower bounds on the two smallest suites; it was not released.

Calibration. The delta sharpened the raw logits (fitted temperature 2.14 → 2.96; raw out-of-domain Brier 0.327, raw locked Brier 0.278). As served, calibration is unchanged. KEV_TEMPERATURE=1.0 gives the raw values.

Previous version: night2-du (2026-09-21), kept at tag night2-du-release

At its release, the recommended Kev. The best accuracy per byte: out of domain 0.797 on the development partition and 0.837 on the locked test, Brier 0.255 on the test, held-out rule pairs 0.77–0.78. This checkpoint is the decision-v7 recipe (trial q35-4b-s23/00-trial-0, seed 2, selected on development accuracy) followed by a 9-minute delta fine-tune (--init_from, lr 2e-5, one epoch) on 1,425 additional records — date-bearing policy cases rendered with explicit day counts, and evidence-free cases with uniform targets — mixed with 2,000 replayed training records. Against the pre-delta checkpoint on the locked test: +1.0 pp [−0.1, +2.1], Brier 0.266 → 0.255, deadline 0.65 → 0.75.

	Kev-4B (Qwen3)	Kev-4B before the delta (v7-base)	Kev-4B, raw logits	Kev-4B as served (T = 2.14)	Jev
in-distribution accuracy (decision-v7 dev, 1,204 records)	0.854	0.877	0.872	0.872	0.845
out-of-domain accuracy (transfer-v4 dev, 764 records)	0.790	0.794	0.797	0.797	0.857
out-of-domain Brier	0.328	0.316	0.299	0.264	0.211
out-of-domain ECE	0.102	0.130	0.122	0.040	0.049
confident errors out of domain (p ≥ 0.9 and wrong)	8.2%	8.2%	6.9%	2.6%	3.7%
coverage at ≤ 5% error (share of decisions automatable)	0.31	0.54	0.54	0.57	0.70
held-out policy structures, both siblings correct	0.73	0.78	0.78	0.78	0.86
unknowable items answered at ≥ 0.9 (lower is better; transfer-v9)	0.44	0.19	0.00	0.00	0.09
locked test, out-of-domain accuracy / Brier	0.806 / 0.294	0.832 / 0.266	0.837 / 0.255	–	–
locked test, in-distribution accuracy	0.856	0.870	0.871	–	–

Per-source out-of-domain accuracy (Kev-4B / Jev): QNLI 0.91 / 0.93, SciQ 0.97 / 0.99, TweetEval-offensive 0.74 / 0.81, PAWS 0.74 / 0.79, MMLU 0.70 / 0.90, Emotion 0.56 / 0.59, deadline (3-level date arithmetic) 0.60 / 0.93 — 0.85 with the date_facts preprocessor (below), (A or B) and C 0.91 / 0.91, (A and B) or not C 0.88 / 0.97, if A then not B else C 1.00 / 0.78.

Calibration is built in. head.pt carries a temperature (T = 2.14) fitted on this checkpoint's in-distribution development rows by minimising negative log-likelihood (scripts/calibrate_checkpoint.py); the pointer head divides its logits by it at inference. Every loader — kev.serve, kev.benchmark, the Space, anyone's harness — gets the calibrated probabilities by default. It never changes an answer: the argmax is identical, so accuracy is the same in both columns; confidences are re-ordered only slightly across questions with different option counts, which is why coverage moves by a point or two. KEV_TEMPERATURE=1.0 restores the raw logits; the raw column is what the training produced. Per-(type, option-count) temperatures were tested and are worse out of domain. The fit uses no out-of-domain or test data.

date_facts preprocessor. Kev, like every Kev before it, cannot subtract dates reliably (the untrained base can; LoRA training erodes it). It can use a stated day count. KEV_DATE_FACTS=1 appends one sentence per pair of absolute dates found in the state ("June 26, 2026 is 8 days before July 4, 2026"); this checkpoint was trained on such renderings, so with it deadline goes from 0.60 to 0.85 and overall out-of-domain accuracy from 0.797 to 0.820. It is preprocessing, reported separately, never folded into the model's own numbers.

What the delta cost. MMLU-Pro fell 0.500 → 0.490 and scienthoon's ECE rose 0.086 → 0.116; coverage at ≤ 5% error was unchanged (0.54 development, 0.67 → 0.68 locked test) and confident errors fell (8.2% → 6.9%). The pre-registered criteria for the delta (PLAN.md, "Tonight's autoresearch") were met for dates and for the unknowable-confidence behaviour; the coverage criterion asked for +5 pp and got 0; the locked read decided promotion.

Newer evaluation columns (transfer-v9 development, Kev-4B / Jev): MMLU-Pro (10-way) 0.490 / 0.840; state buried among unrelated records 0.67 / 0.70; unknowable share at ≥ 0.9 confidence 0.00 / 0.09 (intact controls 0.94).

External suites (same items as their published Jev numbers): SemIf's authored 144 — 0.896 before the delta (live Jev 0.965; SemIf's untrained Qwen3.5-4B 0.813); scienthoon's 900 tickets — queue 0.918, angry 0.790, ECE 0.116 (Jev 0.897, 0.914, 0.105). On ekzhang's 1,000-question MMLU-Pro sample the shipped checkpoint scores 0.468 over all 1,000 questions (8 exceed the state limit and count as wrong; live Jev 0.835 on the same items, ekzhang reports 0.829). On SemIf's pinned third-party selections (evals/external/{wanli,typesafe}-v1): WANLI-256 accuracy 0.695 (live Jev 0.758); TypeSafe-102 equal-case agreement / total-variation distance 0.856 / 0.231 over the 89 rows within the 8,192-token serving context (13 rejected), 0.770 / 0.308 over all 102 with rejected rows scored as wrong (live Jev 0.891 / 0.125; published TypeSafe answers 0.883 / 0.127); plain accuracy on the answered rows 0.843, coverage at <= 5% error 0.02 (Jev 0.892, 0.84). The shipped temperature is fitted in distribution and does not transfer to every workload. On WANLI, a single temperature fitted on the workload's own labelled rows (python -m kev.calibrate, group-disjoint out-of-fold) lowers ECE from 0.166 as shipped to 0.052 (workload T 3.91 against the shipped 2.14). Accuracy is unchanged and coverage at <= 5% error does not improve. On TypeSafe the shipped temperature already fits and refitting does not help (ECE 0.158 as shipped, 0.175 out of fold).

How it was built

• Base model: Qwen3.5-4B-Base, a hybrid of 24 Gated DeltaNet (linear attention) layers and 8 full-attention layers. Because the recurrent layers cannot honour a block-causal mask, questions run as separate causal rows that continue from the shared state (kev/model.py: forward_rows_batch); isolation is exact by construction (together vs alone within 1e-5) and on attention-only models this form is bit-identical to the packed one.
• Recipe: decision-v7, two epochs, LoRA r=16 (attention, MLP and DeltaNet projections), lr 5e-5 — the same data and settings as every other Kev, so the Qwen3 → Qwen3.5 difference is the base (PLAN.md, Qwen3.5 port §10: locked test +7.3 pp [+2.8, +11.7] over Kev-8B).
• Delta: kev.train --init_from jaredpalmer/kev-4b@v7-base --data evals/night2/dates_unknowable.jsonl --replay 2000 --lr 2e-5 --epochs 1. The 1,425 new records are generated (no public dataset): 900 date-bearing policy cases, a third rendered plainly, a third with a relational day-count sentence, a third with a date_facts field; 255 cases with the deciding sentence removed and a uniform soft target over the options, plus their 270 intact controls. Record hashes are in evals/night2/manifest.json; the source checkpoint's hashes are in training_config.json.
• Why a delta and not a retrain: it is a controlled change (one fixed checkpoint, one data addition, 9 minutes), and the results section shows exactly what it moved.

Known limits

• Use Kev-9B when accuracy and calibration matter more than memory: 0.852 vs 0.837 out of domain on the locked test, Brier 0.237 vs 0.255.

• Slow on a Mac. The DeltaNet kernels have no MPS implementation; PyTorch falls back to reference code. A five-question request that takes 0.17 s on the Qwen3 Kev-4B takes 0.78 s here in bf16 on an M5. On CUDA with flash-linear-attention installed it is fast. Use jaredpalmer/kev-4b@qwen3 for low latency on Apple Silicon until an MLX path exists.

• Requires transformers >= 5.17 (the qwen3_5 architecture) and peft >= 0.21.

• Knowledge (MMLU 0.70 vs Jev 0.90; MMLU-Pro 0.490 vs 0.840), TweetEval (0.74 vs 0.81) and noisy-label Emotion (0.56 vs 0.59) are the remaining gap; knowledge is set by the base (the untrained Qwen3.5-4B scores the same).

• Date arithmetic without the preprocessor: deadline 0.60 (Jev 0.93). With KEV_DATE_FACTS=1: 0.85.

• The raw logits are over-confident out of domain; the built-in temperature (T = 2.14) fixes most of it without changing any answer. KEV_TEMPERATURE=1.0 gives the raw values. Coverage at a 5% error budget is 0.54–0.68 against Jev's 0.70.

• 4B bf16 needs ~9 GB of GPU memory for serving; training took 56 min on one H100 (peak 24.6 GB).

Training

Frozen suite evals/v7/decision-v7: 10,000 public records (1,000 per source), 896 policy minimal-pair records over nine template families, 1,680 records from 60 randomly generated rule structures in four rendering styles. Two epochs, LoRA r=16 α=32 on q/k/v/o_proj, gate/up/down_proj, in_proj_qkv/z/a/b, out_proj; 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; option permutation, none-of-the-above insertion, distractors, none minimal pairs on 25% of Choice records. Then the delta described above (one epoch, lr 2e-5, 3,937 records seen, 9 minutes on one H100). No Jev outputs were used for training.

Evaluation protocol

Development partitions select models; the locked test partition is read at most once per candidate (runs/locked/kev-4b-night2-du-ungated/; the pre-delta read is runs/locked/kev-4b-q35/). Every number carries suite hash, code hashes and git commit in result.json. Untrained-base baselines use zero-shot letter logits on the same items (scripts/base_mmlu_probe.py).

Use

uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8008      # KEV_DTYPE=bf16 on a Mac; slow on MPS, see limits
KEV_DATE_FACTS=1 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-4b --port 8008   # + date preprocessing; KEV_TEMPERATURE=1.0 for raw logits

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; the Qwen3.5 base is Apache-2.0; datasets carry their own licenses.