baaderso36/BaaderSo36-DE-Opus4.7-REAP

BaaderSo36-DE-Opus4.7-REAP

A German translation of the BaaderSo36-Opus4.7-REAP reasoning dataset. Each sample preserves the original Claude Opus 4.7 reasoning structure, translated into natural German while keeping format markers (<think>, </think>, Final answer:) and code blocks intact.

Dataset Statistics

• —Total samples: 1,379
• —Source distribution:
• —debug: 511
• —react_advanced: 357
• —humaneval: 161
• —math_hard: 136
• —code_contests: 125
• —math_l5: 67
• —react: 22

Reasoning depth (characters in thinking block)

• —min: 906
• —median: 11,654
• —p90: 22,785
• —p99: 33,827
• —max: 45,503

Long-reasoning buckets:

Sources

The English originals were drawn from publicly available datasets:

• —MATH-500 (Level 5): HuggingFaceH4/MATH-500
• —MATH-Hard (Level 5): lighteval/MATH-Hard
• —HumanEval: openai/openai_humaneval
• —Code Contests: deepmind/code_contests
• —React (basic & advanced): Hand-curated React/Hooks/UI engineering prompts
• —Debugging: Hand-curated cross-language debugging scenarios

Format

Two configs are provided:

messages (default) — chat format

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "<think>...</think>\n\nFinal answer:\n..."}
  ]
}

Load via:

from datasets import load_dataset
ds = load_dataset("baaderso36/BaaderSo36-DE-Opus4.7-REAP", split="train")

split — parser-friendly (9 fields, identical schema to BaaderSo36-EN)

{
  "prompt_id": "math_42",
  "source_dataset": "math_l5",
  "system": "",
  "prompt": "...",
  "thinking": "...",
  "response": "...",
  "think_chars": 4521,
  "resp_chars": 387,
  "model": "claude-opus-4-7"
}

Load via:

from datasets import load_dataset
ds = load_dataset("baaderso36/BaaderSo36-DE-Opus4.7-REAP", "split", split="train")

Translation Method

The original English reasoning was generated by Claude Opus 4.7 (claude-opus-4-7). German translations were produced by:

• —Mistral Small 3.2 (mistralai/Mistral-Small-3.2-24B-Instruct-2506) for the bulk of samples (two-shot prompted)
• —Claude Sonnet 4.5 (claude-sonnet-4-5) for samples where Mistral failed format compliance (rescue pass)

The model field uniformly reports claude-opus-4-7 because the semantic source — the reasoning content — comes from Opus; the translator is transport. Translation provenance metadata was dropped to keep the schema 1:1 identical with the English BaaderSo36 dataset.

Format markers (<think>, </think>, Final answer:) and code blocks were preserved verbatim during translation.

Translation method breakdown (after filtering):

• —llama-rescue: 0 kept / 3 input
• —sonnet-rescue: 465 kept / 477 input
• —two-shot: 914 kept / 967 input

Note: Llama-rescue translations were intentionally dropped due to known quality issues. Only two-shot (Mistral) and sonnet-rescue (Claude Sonnet) translations are retained.

Filter Yield

• —Input: 1,447 translated samples
• —Kept: 1,379 (95%)
• —Dropped: 68

Filter reasons:

• —resp_too_short: 47
• —bad_end_char: 17
• —method_dropped_llama-rescue: 3
• —duplicate_prompt: 1

Intended Use

• —Supervised fine-tuning of language models on German long-form reasoning
• —Combining with English BaaderSo36 to build bilingual reasoning models
• —Research on reasoning preservation across translation

Limitations

• —All reasoning content is translated, not natively generated in German
• —Code blocks and technical terms may retain English where appropriate
• —Translation quality varies between Mistral and Sonnet translations

License

Released under CC-BY-4.0. The underlying generations are produced by Anthropic Claude Opus 4.7 and are subject to Anthropic's Acceptable Use Policy.

Citation

@dataset{baader_so36_de_opus47_reap_2026,
  title = {BaaderSo36-DE-Opus4.7-REAP: A German Translation of Long-Reasoning Distillation Dataset},
  year = {2026},
  author = {baaderso36},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/baaderso36/BaaderSo36-DE-Opus4.7-REAP}
}