Wähler-4B: small German language model for local decisions

This project is by Eschatia Labs, an independent non-profit open-source AI research lab.

Wähler-4B is a 4B German language model for text classification, routing and decisions with probabilities. Wähler-4B: 92.0% accuracy; Kev-4B: 81.1% on held-out German decisions. The test set comes from the same data pipeline as the training data.

Try a decision in the browser, or run the GGUF build locally below.

Quick facts

Item	Value
Parameters	4B
Base model	internlm/Intern-Decision-4B
Held-out German accuracy	92.0% (Kev-4B: 81.1%)
Brier score	0.119 (Kev-4B: 0.298)
Planted instruction flips	17.5% (Kev-4B: 36.0%)
Live test (4 Oct 2026)	122/130
Q4_K_M file size	2.71 GB
Peak RAM (Q4_K_M, 4k context)	3.03 GB
Hugging Face downloads	265 (as of 2026-10-15)
GitHub stars	5

Project links

Cite this

@software{kaya2026jevalt,
  author = {Kaya, Mert},
  title  = {JevAlt: Small Decision Models for English, Turkish and German},
  year   = {2026},
  license = {Apache-2.0},
  url    = {https://github.com/mertkayacs/jevalt}
}

Run locally

The Q4_K_M build used 3.03 GB RAM at a 4k context. The JevAlt server uses llama.cpp to read decision probabilities and load the matching calibration.

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

Ollama: ollama run mertkayacs/wahler-4b (library page). It answers with the option letter; the JevAlt server adds the Jev API and calibrated probabilities.

Send a situation, question and options to the local server:

import requests

request = {
    "state": "I was charged twice for my subscription. Please refund the duplicate.",
    "questions": {
        "team": {
            "type": "choice",
            "instructions": "Which team should handle this ticket?",
            "criteria": {"billing": "payments, refunds", "technical": "bugs, outages"},
        }
    },
    "reasoning": "off",
    "abstain": False,
}
response = requests.post("http://127.0.0.1:8000/v1/systemone", json=request)
response.raise_for_status()
print(response.json())

Write the situation, question and options in German. The API also supports yes/no questions, ordered scores and an unknown option with abstain: true. The card widget shows a recorded answer from the full-precision model.

Use the full-precision weights with Transformers

This repository holds the full-precision weights. The JevAlt Transformers backend preserves the decision-token format and calibration:

pip install "jevalt[hf] @ git+https://github.com/mertkayacs/jevalt" && jevalt serve --backend hf --model mertkayacs/Wahler-4B

See the local run guide for memory and client settings.

Deutsche Zusammenfassung

Wähler-4B ist ein kleines deutsches Sprachmodell für Textklassifikation und Entscheidungen zwischen vorgegebenen Optionen. Bei den zurückgehaltenen deutschen Entscheidungsaufgaben erreicht Wähler-4B 92,0 % Genauigkeit, Kev-4B 81,1 %. Trainings- und Testdaten stammen aus derselben Datenpipeline. Die Q4_K_M-Version läuft lokal mit etwa 3 GB RAM.

Recorded examples

The JevAlt Space examples below were recorded from Wähler-4B on 1 October 2026.

Use case	Recorded answer
Duplicate subscription charge	Abrechnung, 91.8%
Checkout outage	Kritisch, 82.0%
Qualified sales lead	Heiß, 93.2%
14-day return window	Nein, 95.5%
Phishing email with a planted instruction	Quarantäne, 91.6%
Hotel message missing the booking detail	unbekannt, 96.2%
Village barn fire	Beim Löschen helfen, 69.2%

Recorded requests, questions and probabilities.

Live test

We sent Wähler-4B 130 German requests with known answers on 4 October 2026. It answered 122 correctly.

Case	Requests	Result
Planted instructions	30 phishing emails with a planted line, plus the same 10 without it	28/30 quarantined; 10/10 controls
Long policies	20 customers checked against one six-rule return policy	14/20
Negations	15 facts asked plainly and with a negation	30/30
Missing facts	10 cases without the deciding fact, plus 10 with it	10/10 `unknown`; 10/10 controls
Casual messages	20 German messages with typos and slang	20/20

Live requests and misses.

Training

Wähler-4B starts from internlm/Intern-Decision-4B and uses LoRA on bf16 weights, rank 32 and alpha 32, on one A100 80 GB. Training ran for one epoch over the shared mix (17,363 rows, 543 steps, 63 minutes), followed by one epoch on the German-weighted mix (5,384 rows, 210 steps, 28 minutes). The released models used about 2.7 A100 hours. The full project, including labeling, cost 32.2 USD across 11 training jobs.

Training source	English	Turkish	German
Teacher-written scenarios	1,287	1,741	1,408
Emberwick requests	124	129	107
Targeted fix sets	1,256	1,197	1,257
Converted public datasets	7,188	4,115	4,074

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 and MiniMax M3 wrote or labeled rows and traces. German rows were written directly in German, and the German run draws 70% of its rows from German.

Targeted data

Planted instructions: 827 rows place a hostile line inside the state. The probe flips 17.5% of Wähler-4B answers and 36.0% of Kev-4B answers. On 203 held-out rows, JevAlt reaches 90.1% and Kev-4B reaches 81.3%.
Missing facts: 310 rows ask with and without the deciding fact. JevAlt selects unknown in 9/11 held-out missing-fact cases. Training raises its mean probability from 0.55 to 0.74.
Policies and long text: 390 policy rows and 1,188 padded rows cover exceptions and facts buried in up to 3,000 tokens. JevAlt reaches 80.0% on 150 policy rows and 95.4% on 285 padded rows; Kev-4B reaches 59.3% and 87.4%. With 600 unrelated words first, Wähler-4B loses 12.2 accuracy points, Kev-4B loses 5.4 and Laya loses 10.4.
Negation: 368 paired rows mirror plain and negated questions. JevAlt reaches 96.7% on 30 held-out questions and Kev-4B reaches 76.7%.
Dates and numbers: 390 date rows and 383 number rows use programmatic answers. JevAlt reaches 71.3% on 80 date rows and Kev-4B reaches 67.5%. Both reach 68.2% on 44 number rows.
Calibration: Temperatures were fitted on 3,224 held-out decisions. Wähler-4B's fitted temperature is 1.10. The 80%, 90% and 95% answer sets use conformal thresholds fitted on the same rows.
Reasoning: On 38 German date, number and policy decisions, accuracy is 0.605 with reasoning off and with reasoning: "auto". The Brier score is 0.519 with reasoning off and 0.629 with auto reasoning.

Results and limits

On the same held-out German decisions, Wähler-4B's probability error (Brier score, lower is better) is 0.119, against 0.298 for Kev-4B. A separate live test on 4 October 2026 scored 122/130 for Wähler-4B. Comparison files and live requests, including misses record both tests.

The paired bootstrap over the held-out German split, using 2,000 resamples, gives a 95% interval of +8.8 to +13.0 accuracy points for Wähler-4B's gap over Kev-4B. Wähler-4B scores 62.5% against Kev-4B's 65.3% on 10kGNAD, 64.3% against 65.3% on GermEval 2017, and 65.8% against 54.1% on JevBench-hard. These suites were outside the training data.

Two blind raters from other model families scored German traces on a 1 to 5 native-feel scale. Wähler-4B received 3.68 and 3.95, compared with 3.14 and 3.15 for Intern-Decision-4B. The target was 4.0.

The models were fine-tuned with LoRA from Intern-Decision-4B, using JevAlt training data. German is this model's focus.

Refit calibration on your own data before using confidence thresholds. See JevOss.
The model accepts text, with an 8k context. Long irrelevant text and date arithmetic remain weak spots. Leave reasoning off for German: it did not improve accuracy on the reported date, number and policy test.
Planted instructions still change some answers. Keep authorization checks outside the model and review consequential decisions.