Kev-9B

Kev-9B 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, 45.4M trainable parameters) plus a pointer head on Qwen/Qwen3.5-9B-Base (revision 68c46c4b), serving TypeSafe's public /v1/systemone contract.

The most accurate Kev. Out of domain it scores 0.822 on the development partition and 0.852 on the locked test (Jev: 0.857 on the development items), with the lowest Brier of any Kev (0.237 on the test) and held-out rule pairs at 0.81–0.83. This checkpoint is the decision-v7 recipe (trial q35-9b/01-trial-1, seed 1, selected on development accuracy) followed by a 15-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: +1.8 pp [+0.8, +2.9], Brier 0.243 → 0.237, deadline 0.72 → 0.88.

• Hub: jaredpalmer/kev-9b (this repo; trial night2-9b-du/00-trial-0). The pre-delta checkpoint is at revision v7-base.
• Demo: huggingface.co/spaces/jaredpalmer/kev runs Kev-4B and Kev-0.8B on ZeroGPU with the same encoder and API code as kev.serve.
• Code, suites, every trial with hashes and paired bootstraps: github.com/jaredpalmer/kev — PLAN_Qwen35.md (the port and this experiment), PLAN.md, runs/leaderboard.md

Results (same frozen items for every row)

Per-source out-of-domain accuracy (Kev-9B / Jev): QNLI 0.93 / 0.93, SciQ 0.96 / 0.99, TweetEval-offensive 0.78 / 0.81, PAWS 0.76 / 0.79, MMLU 0.74 / 0.90, Emotion 0.60 / 0.59, deadline (3-level date arithmetic) 0.80 / 0.93 — 0.90 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 0.91 / 0.78.

Calibration is built in. head.pt carries a temperature (T = 2.30) 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 0.80 to 0.90 and overall out-of-domain accuracy from 0.822 to 0.828. It is preprocessing, reported separately, never folded into the model's own numbers.

What the delta cost. Coverage at ≤ 5% error fell (0.53 → 0.47 on development; 0.66 → 0.62 on the locked test), confident errors rose (7.5% → 8.7% raw), MMLU-Pro fell 0.545 → 0.515, and scienthoon's ECE rose 0.082 → 0.113. The pre-registered criteria for the delta (PLAN.md, "Tonight's autoresearch") were met for dates and for the unknowable-confidence behaviour and not met for coverage; the locked read decided promotion.

Newer evaluation columns (transfer-v9 development, Kev-9B / Jev): MMLU-Pro (10-way) 0.515 / 0.840; state buried among unrelated records 0.74 / 0.70; unknowable share at ≥ 0.9 confidence 0.00 / 0.09 (intact controls 0.95).

External suites (same items as their published Jev numbers): SemIf's authored 144 — 0.917 before the delta (live Jev 0.965; SemIf's untrained Qwen3.5-4B 0.813); scienthoon's 900 tickets — queue 0.952, angry 0.900, ECE 0.113 (Jev 0.897, 0.914, 0.105). On ekzhang's 1,000-question MMLU-Pro sample the pre-delta checkpoint scores 0.511 (Jev 0.829).

How it was built

• Base model: Qwen3.5-9B-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_Qwen35.md §10: locked test +7.3 pp [+2.8, +11.7] over Kev-8B).
• Delta: kev.train --init_from jaredpalmer/kev-9b@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, 15 minutes), and the results section shows exactly what it moved.

Known limits

• Slow on a Mac. The DeltaNet kernels have no MPS implementation; PyTorch falls back to reference code. A five-question request that takes 0.3 s on Kev-8B takes about 2 s here in bf16 on an M5. On CUDA with flash-linear-attention installed it is fast. Use Kev-8B (Qwen3, jaredpalmer/kev-8b) 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 0.74 vs Jev 0.90; MMLU-Pro 0.515 vs 0.840) is the remaining gap and is set by the base: the untrained Qwen3.5-9B scores the same, and a Kev on the 35B-A3B MoE did not move MMLU-Pro either (PLAN.md, night-2 results).
• Date arithmetic without the preprocessor: deadline 0.80 (Jev 0.93). With KEV_DATE_FACTS=1: 0.90.
• The raw logits are over-confident out of domain; the built-in temperature (T = 2.30) fixes most of it without changing any answer. KEV_TEMPERATURE=1.0 gives the raw values. Coverage at a 5% error budget is 0.47–0.62 against Jev's 0.70.
• 9B bf16 needs ~19 GB of GPU memory for serving; training took 91 min on one H100 (peak 39.5 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, 15 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-9b-night2-du-ungated/; the pre-delta read is runs/locked/kev-9b-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-9b --port 8008      # KEV_DTYPE=bf16 on a 32 GB Mac; slow on MPS, see limits
KEV_DATE_FACTS=1 uv run --extra serve python -m kev.serve --run jaredpalmer/kev-9b --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.