🇩🇪 German R1 is a reasoning model almost equivalent to OpenAI‘s o3 or DeepSeek‘s R1 - but it thinks in German!
🇩🇪 German R1 is so efficient that is was build without any government funding.
🇩🇪 German R1 was only trained with legally imported H100s in less than five minutes.
Please do NOT take this too serious.

Context

See Linkedin: https://www.linkedin.com/posts/activity-7294337496023269376-ZBkI

Usage

You can run the model as follows:

python

from transformers import pipeline, set_seed
import json

set_seed(42)
pipe = pipeline("text-generation", "malteos/german-r1")

# from gsm8k test set
question = "James beschließt, 3-mal pro Woche 3 Sprints zu laufen.  Er läuft 60 Meter pro Sprint.  Wie viele Meter läuft er insgesamt pro Woche?"
expected_answer = "540"

# xml reasoning and answer format
system_prompt = """
Antworte auf deutsch und in diesem Format:
<reasoning>
...
</reasoning>
<answer>
...
</answer>
"""

messages = [
    {
        "role": "system",
        "content": system_prompt,
    },
    {"role": "user", "content": dataset["question"][3]},
]
response = pipe(messages, max_new_tokens=256)

print(json.dumps(response, indent=4, ensure_ascii=False))

The output should look like this (correct answer: 540)

json

[
    {
        "generated_text": [
            {
                "role": "system",
                "content": "\nAntworte auf deutsch und in diesem Format:\n<reasoning>\n...\n</reasoning>\n<answer>\n...\n</answer>\n"
            },
            {
                "role": "user",
                "content": "James beschließt, 3-mal pro Woche 3 Sprints zu laufen.  Er läuft 60 Meter pro Sprint.  Wie viele Meter läuft er insgesamt pro Woche?"
            },
            {
                "role": "assistant",
                "content": "<reasoning>\nJames läuft 3 Sprints pro Tag, wobei jeder Sprint 60 Meter beträgt. Deshalb läuft er pro Tag 3 Sprints * 60 Meter/Sprint = 180 Meter. James läuft 3 Mal pro Woche, also läuft er insgesamt 180 Meter * 3 = 540 Meter pro Woche.\n</reasoning>\n<answer>\n540\n</answer>\n"
            }
        ]
    }
]

Training details

Base model: Qwen2.5-3B-Instruct
Implementation: based on GRPO demo by will brown + language identification reward
Training data: German subset of openGPT-X/gsm8kx (machine translated from the English gsm8k)

License

Qwen research