language:
- de
license:
- cc0-1.0
license_details: >-
Corpus text is sourced from DraCor's GerDraCor export (CC0). Underlying works
(Schiller, d. 1805) are public domain. Code/docs are MIT (see LICENSE).
task_categories:
- text-generation
pretty_name: tiny_schiller
size_categories:
- 1M<n<10M
tiny_schiller
A small (~2 MB) German-language analogue to Karpathy's tiny_shakespeare — 11 of Friedrich Schiller's dramatic works, cleaned and tokenised for tutorial-scale language models.
For a compact, agent-friendly summary (file inventory, load patterns, licensing), see DATA_CARD.md.
"Das Leben ist nur ein Moment, der Tod ist auch nur einer." — Friedrich Schiller
Corpus
~2.07 MB · 11 works · 2,019,857 characters · sourced from DraCor / GerDraCor (CC0). See LICENSING.md for details.
| Tokenizer | Tokens | chars/token |
|---|---|---|
| character-level | 2,019,857 | 1.00 |
| GPT-2 BPE | 854,611 | 2.36 |
cl100k_base |
642,593 | 3.14 |
Use character-level for teaching-scale models (88-token vocab, no tokenizer needed). Use cl100k over GPT-2 when sequence length matters — German umlauts and compounds tokenise 25% more efficiently.
Works
- Die Räuber
- Die Verschwörung des Fiesco zu Genua
- Kabale und Liebe
- Don Carlos, Infant von Spanien
- Wallensteins Lager
- Maria Stuart
- Die Jungfrau von Orleans
- Die Braut von Messina oder Die feindlichen Brüder
- Wilhelm Tell
- Die Piccolomini
- Wallensteins Tod
Quick Start — nanoGPT
python schiller_char/prepare.py # char-level, 88-vocab
python schiller_bpe/prepare.py # GPT-2 BPE, 50k vocab
python schiller_cl100k/prepare.py # cl100k, 100k vocab
HuggingFace Datasets
from datasets import load_dataset
ds = load_dataset("mrkschtr/tiny_schiller")
print(ds["train"][0]["title"])
print(ds["train"][0]["text"][:200])
9 works in train, 2 in test (Wilhelm Tell, Die Braut von Messina). Each row is one complete work with title and text fields.
Instruction & Character Datasets
Pre-built instruction-format parquet files are in data/ on the Hub.
General dialogue style — 7,607 examples teaching Schiller's dramatic register:
ds = load_dataset("mrkschtr/tiny_schiller", data_files="data/instruct.parquet", split="train")
print(ds[0]["prompt"])
print(ds[0]["completion"])
Per-character — fine-tune a model to respond as a specific character:
# 330 examples as Wallenstein · 325 as Carlos · 313 as Fiesco
# 237 as Marquis · 195 as Ferdinand · 194 as Königin · ...
ds = load_dataset("mrkschtr/tiny_schiller", data_files="data/char_WALLENSTEIN.parquet", split="train")
Rebuild locally (generates data/instruct.parquet + 89 data/char_*.parquet files by default):
python scripts/build_instruct.py # all characters
python scripts/build_instruct.py --list-characters # show available characters + turn counts
python scripts/build_instruct.py --character KARL # single character only
Fine-tuning small LLMs
pip install transformers trl datasets accelerate
python examples/finetune_sft.py --model TinyLlama/TinyLlama-1.1B-Chat-v1.0
Default context window is 2048 tokens. Match your model with --context_length:
python examples/finetune_sft.py --model microsoft/Phi-3-mini-4k-instruct --context_length 4096
python examples/finetune_sft.py --model Qwen/Qwen2.5-0.5B --context_length 4096
Tested: TinyLlama 1.1B · Phi-3 Mini 3.8B · Llama 3.2 1B/3B · Qwen2.5 0.5B–3B.
License
Text: public domain (Schiller died 1805) and sourced from DraCor / GerDraCor under CC0. See LICENSING.md for details.
Citation
@misc{schutera2023tinyschiller,
author = {Schutera, Mark},
title = {tiny\_schiller: a small German Schiller corpus for small language models},
year = {2023},
howpublished = {\url{https://github.com/schutera/tiny_schiller}},
note = {Source texts: DraCor / GerDraCor (CC0) and public domain.}
}