Mirrored from autotrust/JEV-9B
All models by autotrust →Fast, calibrated System 1 decisions that are indistinguishable from the closed TypeSafe Jev 1.13 by KL, and deliberate System 2 generation and reasoning from an untouched Qwen3.5-9B — one set of weights, one vLLM engine, routed per request. The fastest model of the family: it answers a single decision in about a third of the time the hosted API takes. Its successor, autotrust/JEV-27B, is closer to Jev, transfers better to unseen tasks and has a stronger System 2.
JEV-9B now takes images. Every step below is one System 1 decision: camera image or screenshot in, a probability
for every action out, in a single forward pass (about 0.2 s on one GPU). Run it with bash vl/serve.sh (see
Images: quick start).
Robot arm: pick and place from camera images. The arm sees a top camera image; at every step System 1 answers two questions (is the target left or right of the gripper, above or below it), and the arm moves accordingly, halving its step whenever an answer flips. It grasps the cube, carries it and drops it in the tray (MuJoCo simulation).
On 20 random scenes it completed the task 10 times; every cube it grasped ended in the tray, and every miss was a grasp 3–5 cm off target. About 165 ms per decision. Asking it to choose one of 8 motor commands directly did not work: this model is a fast visual judge, not an end-to-end controller.
Computer use: screenshot → which element to click. A real browser (headless Chromium). Every clickable element gets a numbered box; System 1 picks the next click (or "the task is complete"), the browser clicks it, and the loop repeats.
95% of 60 random multi-step tasks completed (shop, settings, mail; 3–7 clicks each), about 0.2 s per click. The colour swatches and switches carry no text, so those clicks are decided from the screenshot alone. With the numbered boxes only (no element text) it completed 37%. The failures skipped a step (the colour) and then checked out an empty cart.
Code for both demos: vl/demos/ [blocked]. Image judging, briefly: VL-RewardBench 74.3%, AgentRewardBench AUROC 0.91,
zero-shot short-video recommendation from covers AUC 0.72 (details in reports/vl/ [blocked]).
How it works: JEV-9B's language weights are bit-identical to Qwen3.5-9B's, so vl/serve.sh serves the unmodified
multimodal Qwen3.5-9B with JEV-9B's System 1 adapter (vl/adapter_vllm, the same weights with the layer names moved).
Text decisions match the text-only model (300 test decisions: largest probability difference 0.011). System 2 also reads
images. Keep --max-num-seqs 8 (set in serve.sh); decisions over images are zero-shot.
Integrated System 1 + System 2, first generation. JEV-9B is AutoTrust's first open model to serve both modes of
thinking from a single set of weights; the second generation is
autotrust/JEV-27B. System 1 answers typed questions (noul yes/no ·
choice over 2–16 options · score on a 0–5 scale) in one forward pass and returns a calibrated probability
distribution. System 2 is ordinary text generation with step-by-step reasoning (thinking mode). Both run on the
same backbone in the same engine, and a request chooses its system.
Blocks of Experts recipe. Rather than fine-tuning one monolithic model, the Blocks of Experts (BoE) recipe keeps a strong pretrained model as a frozen expert block and adds a small, detachable expert block trained for one capability. In JEV-9B the System 2 block is Qwen3.5-9B, bit-identical to the release; the System 1 block is 40.2 M trained parameters (0.5 % of the backbone), trained in ≈ 3 hours on one B200. Because the blocks stay separate, adding System 1 costs System 2 nothing: HumanEval is 70.7 % before and after, with all 164 completions byte-identical. Folding the same block into the backbone instead would have cost 9 points (61.6 %).
Indistinguishable from the closed original on System 1, by KL. On the 25,376 held-out questions (53 domains) whose labels are TypeSafe Jev 1.13's own output distributions, the mean KL divergence is ≈ 0.019 (0 = identical). An observer who sees sampled decisions gains on average 0.019 nats of evidence per decision about which model produced it, so it takes about 54 decisions to gather a single nat. The fidelity extends to the teacher's mistakes (see System 1: indistinguishable from TypeSafe Jev 1.13). Among the open Jev reproductions we could find, only the JEV models publish this distribution-level measure (see How JEV-9B compares with other open Jev reproductions).
Faster than the hosted API. On one B200, a single decision takes a median ≈ 90 ms, against 238–301 ms measured independently for the hosted TypeSafe Jev 1.13 API, and one GPU sustains about 15× the decisions per second an independent benchmark achieved against that API (see Speed).
The fast member of the family; JEV-27B is the closer one. Same recipe, same API: JEV-9B is 2.6× faster than JEV-27B on the same benchmark and its weights are a third of the size (18 GB vs 54 GB). JEV-27B lowers mean KL to Jev's distributions from ≈ 0.019 to ≈ 0.017, more than halves KL on unseen task families (0.234 → 0.104), keeps 96 % instead of 90 % of the teacher's accuracy on an independent 16-option benchmark, and scores 78.0 % instead of 70.7 % on HumanEval (see JEV-9B vs JEV-27B).
Two models, two organisations. TypeSafe Jev 1.13 is the hosted, closed-source model made by TypeSafe AI; it is the teacher whose published output distributions this model was trained on. autotrust/JEV-9B (this repository) is an independent open-weights student built by AutoTrust AI from the Apache-2.0 corpus
SargeDev/jev-distill-corpus-v3. It is not affiliated with, endorsed by, or a product of TypeSafe AI, and shares no weights or code with it.
System 1 numbers are on the held-out test_set_30k of jev-distill-corpus-v3. Its 29,955 rows come from three
sources: 25,376 rows labelled with TypeSafe Jev 1.13's own output distributions (yuri_v3), 2,319 Open-Jev rows with
programmatic ground-truth labels (openjev_v2), and 2,260 placeholder rows (yuri_v1). Rows marked Jev-labelled use
only the first group.
JEV-9B was AutoTrust's first integrated System 1 + System 2 model. autotrust/JEV-27B, the second generation, uses the same recipe, code, hyper-parameters, API and two-block packaging; only the backbone and memory settings changed. Both are evaluated on the same held-out test set and the same independent benchmark.
JEV family benchmark highlights (chart from the JEV-27B repository; light bars = JEV-9B). A · KL to Jev by question type · B · accuracy as % of Jev on an independent benchmark · C · System 2 unchanged · D · speed vs the hosted API · click to enlarge
On the fresh Hacker News, V2EX and community examples (illustrations, not a benchmark), JEV-9B got 92 of 96 decisions right against 95 of 96 for JEV-27B. The difference is on the harder tasks: JEV-9B misses a TypeScript port that breaks a "branded, range-checked integer" rule (0.33; JEV-27B 0.93) and flags a CEO wire-transfer fraud message with less confidence (0.56; JEV-27B 0.84).
Which to pick. For routing, moderation, topic triage and short option lists, JEV-9B gives nearly the same answers 2.6× faster (14,400 benchmark decisions in 42 s vs 110 s on one B200) with a third of the weight memory. For long option lists (more than about 8), unfamiliar task families, code-rule checks, fraud screening, or when the System 2 path matters, use autotrust/JEV-27B.
Dozens of open reproductions of TypeSafe Jev appeared within weeks of its launch; the community Jev Decision Index (formerly the Jev Reproductions Tracker) evaluates 55 of them. Most are trained on human or programmatic gold labels, or on their own synthetic data, so they aim to match or beat Jev's accuracy rather than reproduce its probabilities. "Closest to Jev" therefore depends on how closeness is measured:
On the evidence published today, the two JEV models are the closest open models to TypeSafe Jev 1.13 at the level of output distributions, with JEV-9B second to JEV-27B. On the independent benchmark JEV-9B is level with the best of the other models measured there, not ahead of them. It has not yet been run on the Decision Index or JevBench, where AutoJev-27B scores within about one point of Jev and decider-4b v2 edges ahead of it, so we do not claim it is the closest by every measure. Note that some reproductions report beating Jev on their own test sets (AutoJev-27B reports 84.60 % against Jev's 82.79 %); that is a different goal from reproducing Jev's behaviour.
Not to be confused with AutoJev-27B (denis-pplx/autojev-27b), an unrelated Qwen3.8-27B decision model trained with
full-weight SFT on its own data.
TypeSafe does not publish Jev's size or hardware; it reports 70–500 ms end to end. Independent measurements, and ours:
So JEV-9B answers a single decision in roughly a third of the time (JEV-27B in roughly half), and one GPU sustains about 15× (JEV-27B: about 6×) the throughput the benchmark's author achieved against the hosted API. Read these with the caveats: our latencies are measured on the serving host with no network hop, while the hosted numbers include internet, TLS and queueing; hosted throughput depends on client concurrency and the API's rate limits; Jev's latency is roughly flat in the number of questions per request, so bundling questions narrows the throughput gap; and our figures are self-reported while Jev's come from third parties. The two throughput runs use the same benchmark but not an identical call set (ours stops at 16 options).
What the number means. KL(Jev ‖ model) is the expected log-likelihood ratio, per sampled decision, between TypeSafe Jev 1.13 and the student when the decision comes from Jev. On the 25,376 held-out rows whose targets are Jev's own output distributions, the mean is ≈ 0.019 nats (computed from the per-slice values below, which are published to three decimals): one decision carries almost no evidence about which of the two models produced it, and an observer needs about 1 / KL ≈ 54 independent decisions to accumulate one nat (a likelihood ratio of about e ≈ 2.7 : 1).
For scale, Jev is not deterministic itself: an independent study found it changes its answer on 4.3 % of repeated, identical 64-option calls, and it returns probabilities rounded to two decimals, which is the resolution of the targets used here.
JEV-27B reaches ≈ 0.017 (≈ 60 decisions per nat) on the same rows.
The other test rows are not labelled by Jev and are not part of this claim: Open-Jev rows carry programmatic ground
truth (in-distribution KL 0.004 for noul, 0.176 for choice; 0.234 on the OOD split of unseen task families), and
the yuri_v1 rows carry placeholder labels. No Jev-labelled out-of-distribution set exists in the corpus, so the claim
is established on the 53 training domains; outside them, the independent benchmark with human labels (90–97 % of Jev's
accuracy) is the best available evidence.
Fidelity includes the teacher's mistakes. On a poker spot where a solver always checks, TypeSafe Jev 1.13 shoves with 0.62 in a published test; JEV-9B shoves too, with 0.70 (JEV-27B 0.63). A faithful copy of System 1 is also a faithful copy of its blind spots. At 9 B the student also adds some of its own: on an independent benchmark 11.5 % of its 16-option answers change when only the option order changes, against 7.0 % for the teacher (JEV-27B 7.4 %).
Why separate blocks rather than one merged fine-tune. Folding the System 1 LoRA into the backbone would let a
single weight set serve both heads, but it costs generation quality: the merged backbone with the original lm_head
scores 61.6 % (101/164) on HumanEval against 70.7 % for the base, a 9-point drop, even though prose perplexity barely
moves (3.15 → 3.30). Keeping the backbone pristine and applying the System 1 block only on the decision path removes
that trade-off. For decision serving the adapter is merged in memory at start-up, so decision latency matches a
merged bundle.
Why the recipe is this efficient.
lm_head rows for the verbalizer tokens (false/true, 0–5, A–P), so at step 0 its output equals the
pretrained model's zero-shot restricted next-token distribution (verified to |Δp| < 1e-5; measured 8.6e-07). Before
seeing a single label it already agrees with the test targets on 53 % of choice questions with noul AUROC 0.82;
distillation takes it to 90 % / 0.996.choice options is 3.9 %; the same backbone before distillation flips 38 % of the time.score) gives fitted temperatures of 1.002 / 0.984 / 1.012 and ECE 0.0007 with no post-hoc correction.model_path and memory settings changed (the head-initialisation identity holds
there too, 4.5e-07). Going from 9 B to 27 B lowers KL to Jev from ≈ 0.019 to ≈ 0.017, halves OOD KL
(0.234 → 0.104), and raises the System 2 path from 70.7 % to 78.0 % on HumanEval.gazelle93/decision-models-under-pressure (published
25 Sep 2026) asks decision models to pick the right label for real texts from CLINC-150, MTOP, GoEmotions, DBpedia and
financial tweets under three kinds of pressure: more options, near-miss options, and shuffled option order. The labels
are human gold labels, none of this data is in our training set, and TypeSafe Jev 1.13's results are published with it.
We re-ran the same items with autotrust/JEV-9B and autotrust/JEV-27B, up to our 16-option limit.
On data it was never trained on, JEV-9B reaches 97 % of the teacher's accuracy with 2 and 4 options, 94 % with 8 and 90 % with 16: it falls behind faster than JEV-27B (96–98 %) as the option list grows, loses a little more on near-miss options, and is more sensitive to option order than the teacher. Our run follows the benchmark's published method (gold plus the first K−1 distractors of a pool, shuffled per item); the orderings are seeded differently, so compare aggregates, not individual items.
Expected answers were written by hand before the model was run. These are illustrations (≈ 110 decisions), not a benchmark.
Where it failed or wavered:
const port: number = Number(process.env.PORT) against
the rule "a port must be a branded, range-checked integer" (0.33); JEV-27B flags it (0.93).Per-example outputs and the benchmark aggregates are in reports/realworld_9b.json (the HN and V2EX inputs came from
their public APIs on 25 September 2026).
One vLLM engine serves both systems from the same pristine weights. Ordinary requests go through the base
lm_head (System 2, exactly Qwen3.5-9B); requests addressed to the LoRA module jev-decision go through the decision
head (System 1). adapter_vllm/ contains the backbone LoRA plus the 24-slot decision head re-expressed as an lm_head
LoRA (only the 24 verbalizer rows change), so a typed decision is a single prefill step with max_tokens=1,
constrained to the option tokens and read back as log-probabilities.
--logprobs-mode processed_logprobs is required: it makes the returned log-probabilities respect allowed_token_ids.
--max-model-len 4096 is sized for decisions; raise it (for example to 16384) if System 2 requests will think at
length. Add --enable-prefix-caching --mamba-cache-mode align if you ask many questions about the same state.
Set "enable_thinking": true for deliberate, step-by-step reasoning. This path is Qwen3.5-9B unchanged; see the
Qwen3.5-9B model card for its reasoning benchmarks and recommended sampling
settings.
requests + two small JSON files)Values can differ in the third decimal between runs: vLLM computes in bf16 and results depend slightly on which requests are batched together. The head bias and the temperature are applied client-side; the log-softmax normaliser that vLLM applies cancels out, so the result is exactly the decision head's calibrated distribution.
Because both systems live in one engine, a common pattern is to let System 1 answer when it is confident and hand the
rest to System 2. This is a usage pattern, not a configuration we have benchmarked; pick the threshold on your own
validation data, and serve with a --max-model-len large enough for the reasoning budget.
Mixed batches work too: pass a per-request lora_request list (None for System 2, decision for System 1) and both
systems are served in the same generate call.
Notes:
[kind] before [state], so
only questions of the same kind share a prefix. On 293 real states × 7.7 yes/no questions each (≈ 480-token states),
prefix caching served 19.6 % of prompt tokens from cache (+14–20 % throughput) with identical outputs.qwen3_5) support, LoRA on lm_head, --logprobs-mode and
allowed_token_ids; tested with a vLLM development build from September 2026. Start-up takes 3–8 minutes
(CUDA-graph capture with LoRA enabled).The template is tokenised as one string; the last token's final-norm hidden state goes through a
linear fp32 head H → 24 slots (noul → slots 0–1, score → 2–7, choice → 8–23). Inactive slots are masked,
a per-kind temperature is applied, and a softmax yields the distribution aligned with your options. One prefill
pass, no decoding.
transformers + peftoptions are validated: noul must be ["false","true"], score must be ["0".."5"], choice takes 2–16
free-text options. Note that merge_and_unload() above changes the backbone for the rest of the process; to keep both
systems in one process, leave the adapter unmerged and run System 2 inside with model.disable_adapter():.
test_set_30k, temperature applied)test_set_30k, temperature applied)OOD split (13,058 Open-Jev rows from task families not in training, programmatic labels): KL 0.234, top-1 0.918, noul
AUROC 0.989; choice KL 0.351 / top-1 0.837 (game-state decisions are the hardest slice).
Choice option-permutation consistency (1,000 rows × 4 random permutations): mean max |Δp| 0.024, p90 0.055, top-1 flip rate 3.9 %.
Validation KL on a fixed 4 k-row subset; test_set_30k metrics after calibration.
Annealing matters: the v0.7.0 cool-down (500 steps, lr ×0.9 → ×0.02 from step 2000) lowered KL by 25 %; continuing on the unseen remainder of the epoch with the learning rate decayed to ≈ 0.07× (v0.8.0) lowered it by another 24 % and added 1.4 points of choice agreement. JEV-27B folds this into a single cosine schedule.
choice) are measured against
programmatic ground truth, and on the independent benchmark the student reaches 90–97 % of Jev's accuracy, not 100 %.choice distributions are soft (median top-1
probability 0.70). On the 14 % of rows where the teacher's top two options are within 0.1 of each other, argmax
agreement is near chance for any faithful mimic (0.46 where the gap is < 0.05). On teacher-decisive rows agreement
is 0.954, and the student's argmax captures 97.7 % of the teacher probability mass a perfect mimic could (0.693 vs
0.709).noul and score accept only their canonical options; choice accepts 2–16 options. Inputs
longer than 1,024 tokens are truncated (state only, head 60 % / tail 40 %) at serving unless you raise the limit.yuri_v1 memory-relevance stream is 100 % exact-uniform [0.5, 0.5] —
those rows teach nothing about relevance. The model outputs ≈ 0.5 on them by design; do not use it for
memory-relevance scoring without further training.Weights: Apache-2.0 (base model Qwen/Qwen3.5-9B is Apache-2.0; training corpus
SargeDev/jev-distill-corpus-v3 is Apache-2.0, its openjev_v2 stream additionally CC0). The System One framing and
the noul / choice / score primitives originate with TypeSafe AI's Jev; autotrust/JEV-9B is an independent
student model trained on public data and shares no weights, code or affiliation with TypeSafe AI.
hf download autotrust/JEV-9B --include "vl/*" --local-dir JEV-9Bbash JEV-9B/vl/serve.sh # downloads Qwen/Qwen3.5-9B (with its vision encoder) and serves both systems on :8000import base64, requests
def image(path): return {"image": "data:image/png;base64," + base64.b64encode(open(path, "rb").read()).decode()}
r = requests.post("http://localhost:8000/v1/decide", json={ "kind": "choice", "state": ["Top camera image:", image("scene.png"), "\nTask: put the red cube in the tray."], "question": "Is the red cube to the left or to the right of the gripper?", "options": ["left", "right"]}).json()print(dict(zip(r["options"], r["probabilities"]))) ┌── System 2 block: lm_head (248,320 × 4096) ───────► text generation and reasoningRequest ─► Router ─► Qwen3.5-9B backbone (frozen, bit-identical to the base) (adapter off; HumanEval 70.7 % = base) per │ request └── + System 1 block: LoRA (40.1 M) + 24-slot head (98 k) ─► calibrated typed decision (adapter on, decision path only) (one prefill pass; KL ≈ 0.019 to Jev)hf download autotrust/JEV-9B --local-dir JEV-9B # ~18 GBvllm serve JEV-9B --served-model-name autotrust/JEV-9B \ --enable-lora --max-lora-rank 32 --lora-modules jev-decision=JEV-9B/adapter_vllm \ --logprobs-mode processed_logprobs --max-model-len 4096curl -s localhost:8000/v1/chat/completions -H 'content-type: application/json' -d '{ "model": "autotrust/JEV-9B", "messages": [{"role": "user", "content": "In one sentence, what is safety stock?"}], "max_tokens": 60, "chat_template_kwargs": {"enable_thinking": false}}'import json, math, requestsfrom huggingface_hub import hf_hub_download
REPO, URL = "autotrust/JEV-9B", "http://localhost:8000"dh = json.load(open(hf_hub_download(REPO, "adapter_vllm/decision_head.json"))) # bias + verbalizer token idsT = json.load(open(hf_hub_download(REPO, "calibration.json")))["per_kind"] # per-kind temperatures
def decide(kind, state, question, options=None): options = {"noul": ["false", "true"], "score": [str(i) for i in range(6)]}.get(kind, options) lines = options if kind != "choice" else [f"{'ABCDEFGHIJKLMNOP'[i]}) {o}" for i, o in enumerate(options)] prompt = f"[kind] {kind}\n[state] {state}\n[question] {question}\n[options]\n" + "\n".join(lines) + "\n[decision]:" s = dh["slots"]["ranges"][kind][0] ids = dh["verbalizer_ids"][s : s + len(options)] # the option tokens of this kind r = requests.post(f"{URL}/v1/completions", json={ "model": "jev-decision", "prompt": prompt, "max_tokens": 1, "temperature": 1.0, "logprobs": len(options), "allowed_token_ids": ids, "add_special_tokens": False, "return_tokens_as_token_ids": True}).json() lp = {int(k.split(":")[1]): v for k, v in r["choices"][0]["logprobs"]["top_logprobs"][0].items()} z = [(lp.get(t, -1e9) + dh["bias"][s + i]) / T[kind] for i, t in enumerate(ids)] # + head bias, / temperature e = [math.exp(x - max(z)) for x in z] return {o: x / sum(e) for o, x in zip(options, e)}
print(decide("choice", "SKU AX-330 stock at 8% of safety level; supplier late twice this quarter.", "Supplier response for this scenario.", ["issue_warning", "renegotiate", "dual_source", "maintain"]))# ≈ {'issue_warning': 0.33, 'renegotiate': 0.14, 'dual_source': 0.53, 'maintain': 0.001}print(decide("noul", "Customer says the parcel arrived damaged and wants their money back.", "Is the customer asking for a refund?"))def solve(state, question, options, threshold=0.90): p = decide("choice", state, question, options) # System 1: one prefill pass best = max(p, key=p.get) if p[best] >= threshold: return {"system": 1, "answer": best, "distribution": p} prompt = (f"{state}\n\nQuestion: {question}\nOptions: " + "; ".join(options) + "\nThink it through, then give exactly one option on the last line.") r = requests.post(f"{URL}/v1/chat/completions", json={ # System 2: same engine, base lm_head "model": "autotrust/JEV-9B", "messages": [{"role": "user", "content": prompt}], "max_tokens": 8192, "chat_template_kwargs": {"enable_thinking": True}}).json() return {"system": 2, "reply": r["choices"][0]["message"]["content"], "system1_distribution": p}from vllm import LLM, SamplingParamsfrom vllm.lora.request import LoRARequest
llm = LLM("JEV-9B", enable_lora=True, max_lora_rank=32, logprobs_mode="processed_logprobs", max_model_len=4096)decision = LoRARequest("jev-decision", 1, "JEV-9B/adapter_vllm")
gen = llm.generate(["..."], SamplingParams(temperature=0.0, max_tokens=256)) # System 2, no LoRAdec = llm.generate([prompt], [SamplingParams(max_tokens=1, temperature=1.0, # System 1 allowed_token_ids=ids, logprobs=len(ids))], lora_request=decision) # then + bias, / T as above[kind] choice[state] SKU AX-330 stock at 8% of safety level; supplier late twice this quarter.[question] Supplier response for this scenario.[options]A) issue_warningB) renegotiateC) dual_sourceD) maintain[decision]:import json, torchfrom huggingface_hub import hf_hub_downloadfrom peft import PeftModelfrom safetensors.torch import load_filefrom transformers import AutoModelForCausalLM, AutoTokenizer
repo = "autotrust/JEV-9B"tok = AutoTokenizer.from_pretrained(repo)base = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda") # == Qwen3.5-9B text model
# --- System 2: the pristine base model, no adapter ----------------------------------------------msgs = [{"role": "user", "content": "In two sentences, what is safety stock?"}]enc = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")out = base.generate(**enc, max_new_tokens=80)print(tok.decode(out[0, enc["input_ids"].shape[1]:], skip_special_tokens=True))
# --- System 1: attach the LoRA adapter (merged here for speed) + the 24-slot head ---------------model = PeftModel.from_pretrained(base, repo, subfolder="adapter").merge_and_unload()head = load_file(hf_hub_download(repo, "head.safetensors"))cfg = json.load(open(hf_hub_download(repo, "judge_config.json")))temp = json.load(open(hf_hub_download(repo, "calibration.json")))["per_kind"]W, b = head["proj.weight"].cuda(), head["proj.bias"].cuda()
def decide(kind, state, question, options): letters = "ABCDEFGHIJKLMNOP" lines = options if kind != "choice" else [f"{letters[i]}) {o}" for i, o in enumerate(options)] text = f"[kind] {kind}\n[state] {state}\n[question] {question}\n[options]\n" + "\n".join(lines) + "\n[decision]:" ids = tok(text, return_tensors="pt", add_special_tokens=False).to("cuda") with torch.no_grad(), torch.autocast("cuda", dtype=torch.bfloat16): h = model.model(**ids).last_hidden_state[0, -1].float() # backbone only, last token z = (W @ h + b) / temp[kind] s, _ = cfg["slots"]["ranges"][kind] p = torch.softmax(z[s : s + len(options)], 0) return dict(zip(options, p.tolist()))
print(decide("noul", "Customer says the parcel arrived damaged and wants their money back.", "Is the customer asking for a refund?", ["false", "true"]))# {'false': 0.009, 'true': 0.991}model-0000{1..5}-of-00005.safetensors Qwen3.5-9B text backbone incl. lm_head — bit-identical to the base model (bf16; GDN A_log / gated-norm weights fp32 as in the original), 17.9 GBmodel.safetensors.index.json · config.jsonadapter/ System 1 LoRA (peft format, r=16, 40.1 M params, 154 MB) — apply only for decisionshead.safetensors 24-slot decision head (fp32): proj.weight [24, 4096], proj.bias [24]judge_config.json slot layout, verbalizer token ids, template version, weights_mode=unmerged, provenancecalibration.json per-kind temperatures (+ fit diagnostics)adapter_vllm/ the same adapter for vLLM: backbone LoRA (zero-padded to r=32) + decision head as an lm_head LoRA, plus decision_head.json (head bias, verbalizer token ids)tokenizer.json · tokenizer_config.json · chat_template.jinjareports/ evaluation reports: test-set evaluation, bundle checks, HumanEval per-problem results, vLLM measurements, real-world tests, training-milestone reviewsvl/ vision: serve.sh (multimodal Qwen3.5-9B + System 1), serve_decide.py (POST /v1/decide), adapter_vllm/ (layer names for the multimodal model), calibration.json, demos/videos/ robot-arm and computer-use demo videosreports/vl/ image evaluations and demo results@misc{autotrust_jev9b_2026, title = {autotrust/JEV-9B: the first integrated System 1 + System 2 open model built with the Blocks of Experts recipe (Qwen3.5-9B; System 1 distilled from TypeSafe Jev 1.13)}, author = {{AutoTrust AI}}, year = {2026}, url = {https://huggingface.co/autotrust/JEV-9B}}