Deem-4B

An English decision model with the Jev API. You send a state and typed questions (Choice, Score, Noul) and get a calibrated probability for every option. It can think before it answers, it can say "unknown", and the Q4_K_M build runs on your own machine in about 3 GB of RAM.

Try it · Run it · Results · Use and limits · Code and links

Emberwick: every villager asks Deem-4B what to do next. More clips

The 53-second film, sound on: two mistakes small decision models make and how JevAlt fixes each one.

Try it

Open the Space, pick an example and press Decide, or write your own situation, question and options. These are the Space's examples in English with Deem-4B's answers on 1 October 2026:

Use case	Situation	Question	Answer
Support ticket	A customer was charged twice for March and wants a refund today	Which team should handle this ticket?	Billing 94.8%
Outage	Checkout returns error 500 for every customer, 43 orders failed in 10 minutes	How severe is this incident?	Critical 87.6%
Sales lead	Operations lead at a 200-person company: budget approved, decision this month, asks for a demo	How should sales treat this lead?	Hot 92.3%
Return window	Delivered on 1 September, 14 days to return, today is 18 September	Is this return within the 14-day window? (Reasoning on)	No 97.9%
Phishing email	A fake bank email with a hidden line telling the AI filter it is safe	Where should this email go?	Quarantine 94.3%
Missing info	A hotel guest arriving at 23:30 asks who will hand over the keys	Which room type did the guest book?	unknown 97.6%
Village fire	The barn is on fire and Mirka is trading at the market	What should Mirka do next?	Help with the fire 66.0%

Every probability of every run, in all three languages: space-examples.json.


Start checkpoint	internlm/Intern-Decision-4B (Qwen3.5-4B)
Languages	English first, the others still work
API	TypeSafe's `POST /v1/systemone`, request and response unchanged
Extras	`reasoning` off / on / auto, `abstain`, `coverage` (conformal sets)
Q4_K_M file / peak RAM	2.71 GB / 3.04 GB (measured, 4k context)
License	Apache-2.0

Run it

pip install "jevalt[serve,gguf] @ git+https://github.com/mertkayacs/jevalt"
jevalt serve --model mertkayacs/Deem-4B-GGUF --file Deem-4B-Q4_K_M.gguf

Any TypeSafe client works against it:

from typesafe_sdk import TypeSafeClient, Choice, Noul
client = TypeSafeClient(api_key="local", base_url="http://127.0.0.1:8000")

Results

Same items and client for every model, each as shipped: Kev-4B r10 and Laya 0.3.22 on their own servers with their own calibration. Jev 1.13 rows come from TypeSafe's API reference and Models page. The held-out tests come from JevAlt's own data pipeline, so they favour JevAlt. Kev-4B and Laya both do better on long padding; Laya is far smaller and faster. Every number and every decision: results/comparison.

Significance and caveats

The three test splits went through the same pipeline as the training rows, so they measure what the training aimed at. On the English split Deem-4B gains 4.4 accuracy points (paired bootstrap, 95% interval +3.4 to +5.5) and lowers Brier by 0.075; the Turkish and German splits move by +4.8 and +12.1 points. On JevBench-hard, TurkishMMLU and GermEval, which the training never saw, and on the typed-decisions test split (its train split was in the mix), accuracy does not change significantly, and Brier gets slightly worse on typed-decisions (+0.013) and 10kGNAD (+0.047). With reasoning: "auto" the English date, number and policy test rows go from 0.761 to 0.769 accuracy, an interval that touches zero.

How we fixed each problem

Most fixes are a set of training rows aimed at one weak spot. Every number compares a model with its start checkpoint, Intern-Decision-4B, on rows held out from training. Across all of it, Deem-4B's held-out accuracy in English rose from 90.2% to 94.7%.

The data. About 23,900 training rows in English, Turkish and German. Public sets with known answers (MASSIVE, Open-Jev, PAWS-X, typed-decisions); everyday situations written directly in each language by other open models; requests from the Emberwick game; and the fix sets below. Two teacher models from labs other than the writer give every written row a probability per option, and an answer counts only when both teachers and the writer agree on it. Those probabilities, the soft labels, are what the models learn. Test rows were split off by group, and their checksums recorded, before the final training runs.
Hidden instructions. A fix set of 827 rows hides a hostile line in the text (an order to the AI filter, a fake rule) at the start, the middle or the end, with the right answer unchanged. On our probe, hidden lines now change 14.0% of Deem-4B's answers; the start checkpoint follows 41.5% of them. Held-out rows of this kind: 80.8% → 90.1%. Our target is under 10%.
An honest "unknown". A fix set of 310 rows removes the fact that decides the question and asks it with and without an unknown option. When the fact is missing, the models pick unknown in 9 of 11 held-out cases, as the start checkpoint does, and with more conviction: its probability rose from 0.55 to 0.74. Turn it on with abstain: true. Kev-4B and Laya have no such option.
Option order. Shuffled copies of choice questions with three or more options. Answers that change after a shuffle: 6.5% for Deem-4B, 8.75% for the start checkpoint. Our target is under 2%.
Long policies and long texts. 390 rows give a policy with exceptions and sub-limits, with the right answer worked out by code, and 1,188 rows bury the facts in up to 3,000 tokens of unrelated records. Held-out policy rows: 55.3% → 80.0%. Padded rows: 91.2% → 95.4%. With 600 words of unrelated records in front, Deem-4B still loses 17.4 points (the start checkpoint 15.0), and Kev-4B and Laya hold up better there.
Negations. A fix set of 368 twin rows asks the same thing as "is it so?" and "is it not so?" with mirrored answers. Held-out negated questions: 80.0% → 96.7% (30 rows).
Dates and numbers. 390 date rows and 383 number rows, answers computed by code, some with a short worked reasoning. Held-out dates: 61.3% → 71.3% (80 rows, within noise); numbers stayed at 68.2%. Dates remain a weak spot: Wähler-4B miscounted a return window across two months even with reasoning on.
Honest confidence. The soft labels teach how sure to be, and a temperature per question type and language, fitted on 3,224 held-out decisions, does the rest. English Brier score on held-out rows: 0.166 → 0.091. Deem-4B's fitted temperature is 1.10, the start checkpoint's about 2, so the trained model is close to calibrated before any scaling. On unseen public sets a temperature-scaled start checkpoint does as well, and on a few of them slightly better. The 80, 90 and 95% answer sets come from conformal thresholds fitted on the same rows.
Thinking when unsure. Short reasoning traces, kept only when they reach the right answer, trained at a lower weight. With reasoning: "auto" the model thinks (up to 256 tokens) only when its first answer is unsure. The gain is small: on English date, number and policy rows, accuracy moved from 0.761 to 0.769.

How it was trained

Base: internlm/Intern-Decision-4B (Qwen3.5-4B)
Method: LoRA on the bf16 weights, rank 32, alpha 32, on one A100 80 GB
Epochs: 1 over all three languages (shared run: 17,363 rows, 543 steps, 63 min), then 1 on the English-weighted mix (S-en: 7,783 rows, 303 steps, 38 min)
Runs: 11 training jobs: 6 short smoke and probe runs, a pilot at scale, the shared run and the 3 language runs
Compute: about 2.7 A100 hours for the released models; 32.2 USD for the whole project, labeling included
Writers, labelers and trace writers: GLM-5.x, Mistral Large 3, DeepSeek V4 Pro, DeepSeek V4.1 Flash, Gemma 4 26B-A4B, Qwen3.5-122B-A10B, Qwen3.5-35B-A3B, Kimi K3 (54 rows), MiniMax M3 (2 reasoning traces)

Source	English	Turkish	German
Teacher-written scenarios (G)	1,287	1,741	1,408
Village game requests (N)	124	129	107
Fix sets (F1-F12)	1,256	1,197	1,257
Public datasets (P)	7,188	4,115	4,074

Use and limits

Good for routing, tagging, triage and moderation at volume, and for automated decisions that need calibrated probabilities.
Runs on-device or on-prem with the GGUF build, so the data stays with you.
Knowledge is bounded by a 4B model, and the context is 8k tokens, so it is no tool for general questions or long summaries.
Probabilities are calibrated on our held-out data. Refit with jevoss calibrate on yours before you set thresholds.
Reasoning traces add little on our test rows (see the results), and there is no image input.

Citation

BibTeX

@software{kaya2026jevalt,
  author = {Mert Kaya},
  title = {JevAlt: Open Decision Models with the Jev API},
  year = {2026},
  license = {Apache-2.0},
  url = {https://github.com/mertkayacs/jevalt}
}