kev-4b-ONNX

This is a Transformers.js-ready ONNX conversion of the original Hugging Face model jaredpalmer/kev-4b. The original model card content follows below.

Usage with Transformers.js

Variants under onnx/: q4f16 (default) and q4 (4-bit MatMulNBits weights, block 32, embedding as GatherBlockQuantized, pointer head in fp32; q4f16 runs the rest of the graph in fp16 for WebGPU). The full-precision export is not published.

Kev is not a text generator and not a plain sequence classifier. One document (the state) and any number of typed questions are packed into a single sequence with five delimiter tokens; a block-causal mask lets every question see the state and itself only, and a pointer head scores each option's </opt> token against its question's <decide> token. The graph takes only input_ids and attention_mask (right padded) and derives segments, positions and the mask from the delimiter ids in-graph. Its logits output has one value per token: read the value at every option's </opt> position and softmax within the question. config.json carries the delimiters, ids and limits under kev.

Transformers.js resolves qwen3 to its text-generation class, which expects a KV-cache graph, so load this graph through the base PreTrainedModel class: unknown or unmapped types take the single-session encoder-only path, which feeds every graph input by name. It logs one warning ("assuming encoder-only architecture"), which is expected.

import { AutoTokenizer, PreTrainedModel, Tensor } from "@huggingface/transformers";

const repo = "onnx-community/kev-4b-ONNX";
const tokenizer = await AutoTokenizer.from_pretrained(repo);
const model = await PreTrainedModel.from_pretrained(repo, { dtype: "q4f16", device: "webgpu" });

// Caller text can never produce a delimiter: <|name|> -> <¦name¦> (kev.model.user_tokens)
const enc = (t) => Array.from(tokenizer(t.replace(/<\|([A-Za-z0-9_]+)\|>/g, "<¦$1¦>"), { add_special_tokens: false }).input_ids.data, Number);
const [STATE, Q, OPT, END, DECIDE] = ["<|fim_prefix|>", "<|fim_middle|>", "<|box_start|>", "<|box_end|>", "<|fim_suffix|>"].map((t) => enc(t)[0]);

// TypeSafe request: choice = criteria object (name -> description or null), noul = yes/no, score = ordered levels
const state = "I was charged twice for the same order and nobody answers my emails. I want my money back now.";
const questions = [
  { instr: "Which team should handle this ticket?", options: ["billing: Charges, refunds, invoices", "technical support: Bugs, outages, login problems", "sales", "account management"] },
  { instr: "Is the customer asking for a refund?", options: ["no", "yes"] },
  { instr: "How positive is the sentiment of this message?", options: ["very negative", "negative", "neutral", "positive", "very positive"] },
];

const tokens = [STATE, ...enc(state)];
const groups = questions.map((q) => {
  const spans = q.options.map((o) => [OPT, ...enc(o), END]);
  const branch = [Q, ...enc(q.instr), ...spans.flat(), DECIDE];
  const base = tokens.length;
  let cursor = 1 + enc(q.instr).length;
  const ends = spans.map((s) => { cursor += s.length; return base + cursor - 1; }); // index of each </opt>
  tokens.push(...branch);
  return ends;
});

const { logits } = await model({
  input_ids: new Tensor("int64", BigInt64Array.from(tokens, BigInt), [1, tokens.length]),
  attention_mask: new Tensor("int64", new BigInt64Array(tokens.length).fill(1n), [1, tokens.length]),
});
const scores = Array.from(logits.to("float32").data);
const softmax = (xs) => { const m = Math.max(...xs); const e = xs.map((x) => Math.exp(x - m)); const s = e.reduce((a, b) => a + b); return e.map((x) => x / s); };
const answers = groups.map((ends) => softmax(ends.map((i) => scores[i])));
// choice: argmax over the criteria keys; noul: answers[i][1] is p(yes); score: expected level = sum(i * p[i])

Limits: the state is cut to 8,192 tokens and each question branch (instruction, options, delimiters) plus the state must fit in 8,192 tokens (training used 384 and 1,024; longer inputs run but are untested). Choice questions take 1 to 255 options. Option descriptions render as name: description; structured state or instructions (objects, lists) render as indented key: value text as in the source kev.api.render.

---

Original model card

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) plus a pointer head on Qwen/Qwen3-4B-Base, serving TypeSafe's public /v1/systemone contract.

The recommended kev. The best 4B checkpoint under a frozen, checksummed protocol after ~40 controlled 4B trials, and the first Kev within seven points of Jev out of domain on the same items. Same recipe run at three seeds: transfer 0.773 / 0.790 / 0.770; this checkpoint is the seed selected on the development partition (never on the locked test).

• 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 rule pairs 0.62 / 0.73 / 0.67 (Jev 0.86); 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 checkpoints.

Locked test, read once (runs/locked/kev-4b-v7-preview-ungated/): 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 0.48 (Kev-8B 0.95, Kev-0.6B 0.97) while picking the return reason correctly (wrong size 0.53; Kev-8B 0.84; Kev-0.6B prefers "none of the above" 0.58). 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.