Upload 4 files

DATA_CARD.md

@@ -0,0 +1,336 @@
+---
+license:
+  - other          # source content: public domain (Schiller, d. 1805)
+  - mit            # pipeline code (this repo)
+language:
+  - de
+multilinguality:
+  - monolingual
+size_categories:
+  - 1M<n<10M
+source_datasets:
+  - original
+task_categories:
+  - text-generation
+  - fill-mask
+task_ids:
+  - language-modeling
+  - masked-language-modeling
+pretty_name: tiny_schiller
+tags:
+  - literature
+  - drama
+  - german
+  - 19th-century
+  - public-domain
+  - schiller
+  - tiny-language-models
+---
+
+# Dataset Card for tiny_schiller
+
+> *"Das Leben ist nur ein Moment, der Tod ist auch nur einer!"* — Friedrich Schiller
+
+## Dataset Description
+
+`tiny_schiller` is a small (~2 MB) plain-text corpus of Friedrich Schiller's
+dramatic works, intended as a German-language analogue to Andrej Karpathy's
+[`tiny_shakespeare`](https://huggingface.co/datasets/tiny_shakespeare). It is
+sized for tutorial-scale language models — character-level RNNs, small
+GPT-style transformers, BPE tokenizer experiments — where the goal is
+reproducibility and pedagogy, not state-of-the-art benchmarking.
+
+The corpus is a single concatenated file containing 13 works
+(10 dramas, the unfinished *Demetrius* scenario, plus *Briefe über Don Carlos*), drawn from Projekt
+Gutenberg-DE.
+
+- **Curated by:** Mark Schutera (`schutera`)
+- **Languages:** German (de), 19th-century editorial orthography
+- **License:** see *Licensing Information* below — dual: source content is
+  public domain; pipeline code is MIT
+- **Repository:** <https://github.com/schutera/tiny_schiller>
+
+### Intended Use
+
+**Tutorial and teaching**
+
+- Teaching language-model fundamentals on a non-English literary corpus
+- Tokenizer experiments (character-level, GPT-2 BPE, `cl100k_base`)
+- Small-scale German text-generation demos and overfitting studies
+- A drop-in replacement for `tiny_shakespeare` in lecture material
+
+**Fine-tuning and domain adaptation**
+
+- Continued pre-training or domain adaptation target for any fine-tuneable
+  causal LM (e.g. GPT-2, LLaMA, Mistral, Phi): the corpus is clean,
+  stylistically consistent, and free of web noise
+- Style transfer experiments: the corpus provides a strong, distinctive
+  register (elevated 19th-century dramatic German) that can be imprinted
+  on a base model through a short fine-tuning run
+- Persona modeling: speaker turns are tagged, so individual characters can
+  be isolated and used as instruction-tuning or RLHF preference data targets
+- Historical-German language adaptation: useful for probing how well a model
+  generalises to pre-1901-reform orthography after fine-tuning
+
+**Hugging Face Transformers integration**
+
+- The corpus loads directly with `datasets.load_dataset` and works with the
+  Transformers `Trainer` / `SFTTrainer` API for causal LM fine-tuning with
+  no preprocessing beyond tokenization
+
+**Evaluation**
+
+- Held-out perplexity reference for German literary prose and drama: the
+  `train_test` split provides a small but stylistically coherent test set
+  that is complementary to modern German benchmarks (GermanQuAD, MLQA, etc.)
+
+### Out-of-Scope
+
+- Production language modeling (size and coverage are deliberately small;
+  fine-tuning a base model on this corpus alone will not produce a
+  general-purpose German assistant)
+- Stylometry or authorship attribution (single author, single source edition)
+- Modern German NLP evaluation (orthography is 19th-century; perplexity
+  scores are not comparable to contemporary German benchmarks)
+- Critical / scholarly use (no annotations, no critical apparatus)
+- Representative German prose: the corpus is overwhelmingly dramatic dialog;
+  it is not a substitute for a balanced German text corpus
+
+## Languages
+
+German (`de`). The text reflects the orthographic conventions of the
+19th-century print editions used as source material — pre-1901 spelling
+reform conventions, German chevron quotation marks (»…«), and editorial
+punctuation that does not match contemporary Duden norms.
+
+## Dataset Structure
+
+### Data Instances
+
+A single `text` field containing the concatenated corpus.
+
+```python
+FeaturesDict({
+    'text': Text(shape=(), dtype=string),
+})
+```
+
+### Data Fields
+
+- `text` *(string)*: full concatenated text of all included works.
+
+### Splits
+
+The default release ships a single split. Two configurations are provided:
+
+| Config         | Split   | Bytes (UTF-8)    |
+|----------------|---------|------------------|
+| `default`      | `train` | 2,216,450 bytes  |
+| `train_test`   | `train` | ~90% of total    |
+| `train_test`   | `test`  | ~10% of total    |
+
+The 90/10 split is contiguous-tail (last 10% of the concatenated stream is
+held out), matching `tiny_shakespeare` convention; it is **not** a per-work
+split and therefore leaks within-work distribution between train and test.
+Treat it as a sanity-check held-out, not a benchmark split.
+
+### Line Organization
+
+After Phase 1b cleaning, the file follows the Karpathy `tiny_shakespeare`
+convention: each speaker turn is preceded by a SPEAKER tag on its own line,
+followed by the spoken text. Stage directions appear inline in parentheses
+or on their own lines, as in the source editions. Act and scene markers
+are preserved as headings.
+
+### Token Counts
+
+| Tokenizer       | Token count   | chars/token |
+|-----------------|---------------|-------------|
+| character-level | 2,168,278     | 1.00        |
+| GPT-2 BPE       | 913,675       | 2.37        |
+| `cl100k_base`   | 681,548       | 3.18        |
+
+### Tokenizer Choice
+
+**Character-level** is the right default for tutorial use. The 97-character
+vocabulary is small enough that an embedding table and output projection add
+negligible parameters to a teaching-scale model. Every codepoint is
+unambiguous — no out-of-vocabulary tokens, no merging surprises — and the
+prepare script is a single `np.frombuffer`. This is what `tiny_shakespeare`
+uses and why this corpus was designed to be a drop-in replacement.
+
+**GPT-2 BPE** (`gpt2` via tiktoken) is the right choice when you want to
+reuse a pre-trained tokenizer from Karpathy's nanoGPT tutorials without
+modification. The 50,257-token vocabulary encodes German with 2.37
+chars/token — noticeably worse than cl100k because GPT-2's merge table was
+learned on English text. Umlauts (ä, ö, ü, Ä, Ö, Ü) and ß appear
+frequently in Schiller and they fall outside the most common ASCII merges,
+so many compound words are split more aggressively than necessary. Token ids
+fit in `uint16` (max id 50,256), so `train.bin` / `val.bin` are compact.
+Ready-to-use files are in `schiller_bpe/`.
+
+**`cl100k_base`** (GPT-4 / Claude-era tokenizer) is the best choice when
+sequence length matters — 682k tokens versus 914k for GPT-2 on the same
+text means 25% more context per fixed-length window. The merge table was
+trained on multilingual data and handles German compounds and umlauts much
+more cleanly (3.18 chars/token). The trade-off: the vocabulary is ~100k
+tokens, so ids require `uint32` rather than `uint16`, and the output
+projection over 100k classes may be disproportionately large for a
+tutorial-scale model. Ready-to-use files are in `schiller_cl100k/`.
+
+**When to choose which:**
+
+| Goal | Tokenizer |
+|------|-----------|
+| Drop-in for nanoGPT / makemore tutorials | character-level |
+| Follow Karpathy's nanoGPT BPE tutorial exactly | GPT-2 BPE |
+| Minimise sequence length / maximise context | `cl100k_base` |
+| Cross-lingual transfer from a GPT-4-style model | `cl100k_base` |
+| Explore tokenizer design for German | compare all three |
+
+## Data Sources
+
+All texts are sourced from [Projekt Gutenberg-DE](https://www.projekt-gutenberg.org/autoren/namen/schiller.html)
+(projekt-gutenberg.org). Note: this is the German Projekt Gutenberg-DE,
+operated by Hille & Partner / *Der Spiegel*, and is **distinct** from the
+US-based Project Gutenberg at gutenberg.org. The two have different reuse
+terms; see *Licensing Information* below.
+
+Per-work source URLs are recorded in `scripts/sources.yaml`. Note: SHA-256 hashes of the original HTML are unavailable — the Projekt Gutenberg-DE site restructured in early 2026 and original chapter HTML files are no longer publicly accessible for re-verification.
+
+Included works:
+
+- Briefe über Don Carlos
+- Aus dem Szenar zum »Demetrius«
+- Demetrius
+- Die Huldigung der Künste
+- Die Jungfrau von Orleans
+- Die Räuber
+- Die Verschwörung des Fiesco zu Genua
+- Don Carlos, Infant von Spanien
+- Kabale und Liebe
+- Maria Stuart
+- Wallenstein (trilogy as published)
+- Wilhelm Tell
+- Die Braut von Messina
+
+## Curation Rationale
+
+`tiny_shakespeare` is a teaching staple because it is small, homogeneous,
+and rich in dialog structure. There is no equally convenient German
+counterpart. Schiller is a natural choice: a single canonical author,
+multiple long dramas with explicit speaker turns, and a body of work
+entirely in the public domain. The goal was a single plain-text artifact
+that drops into existing tutorial pipelines with minimal friction, plus
+a reproducible audit trail (sources, hashes, cleaning script).
+
+## Considerations for Using the Data
+
+The corpus is a single 19th-century male author writing about idealised
+historical and political subjects; it encodes the worldview and
+historiographic assumptions of late-Enlightenment German drama. Models
+trained on it will reflect that distribution and should not be used to
+generate text presented as contemporary or neutral.
+
+- **Orthography:** pre-1901-reform German; downstream evaluation against
+  modern German benchmarks is unfair.
+- **Quotation marks:** German chevrons »…« are preserved; tokenizers may
+  treat them as rare characters.
+- **Dialog-heavy:** overwhelmingly dramatic dialog with speaker tags;
+  not a representative sample of German prose.
+- **Editorial layer:** old non-critical print editions; spelling and
+  punctuation reflect the digitiser's editorial choices, not Schiller's
+  manuscripts.
+- **Size:** ~2 MB / 11 works is a teaching toy. Schiller's poetry,
+  philosophical essays, and historical writing are excluded in v1.
+- **Encoding artifacts:** the cleaning pipeline normalises common HTML
+  boilerplate and mojibake; residual artifacts are tracked in the audit
+  log.
+
+## Licensing Information
+
+This dataset has **two distinct licenses**, one for the source content
+and one for the pipeline that produces it.
+
+### Source content — public domain
+
+Friedrich Schiller died on 9 May 1805. Under the EU `Schutzdauerrichtlinie`
+(life + 70 years) and equivalent rules in the United States and the
+United Kingdom, all of Schiller's works passed into the public domain
+no later than 1 January 1876, and copyright in the original works has
+not been recoverable since. The texts themselves are therefore free of
+authorial copyright.
+
+The print editions used as the source for the Projekt Gutenberg-DE
+digitisations are old non-critical editions; they do not carry separate
+editorial / critical-apparatus copyright that would survive in the body
+text. (Modern critical / annotated editions — e.g. the *Nationalausgabe*
+or Hanser editions — are **not** used here, and their annotations would
+remain in copyright.)
+
+### Pipeline code — MIT
+
+The scripts, configuration, and documentation in this repository are
+licensed under the MIT License (see `LICENSE`). MIT covers only the
+code, not the textual content.
+
+### Source-platform terms — Projekt Gutenberg-DE
+
+Projekt Gutenberg-DE (projekt-gutenberg.org) is an independent compilation
+operated by Hille & Partner / *Der Spiegel*, and applies its own terms
+to redistribution of *its own digitisations* — independent of the
+underlying public-domain status of the works. In paraphrase: the texts
+are made available for **personal, private, non-commercial use**;
+**redistribution of the digitised text in unaltered form requires the
+permission of the editor** (Editor: Gunter Hille). Schools and individuals
+may use the texts for personal, classroom, and research purposes.
+
+This dataset takes the conservative position that:
+
+1. The underlying works are public domain and may be re-derived from
+   any source edition, and
+2. As a courtesy and to honour the source-platform terms, attribution
+   to Projekt Gutenberg-DE is preserved at the per-work level
+   (`scripts/sources.yaml`), and the corpus is released for research
+   and educational use.
+
+For full details, exact wording, and the redistribution edge cases,
+see [`LICENSING.md`](LICENSING.md).
+
+## Citation
+
+```bibtex
+@misc{schutera2023tinyschiller,
+  author       = {Schutera, Mark},
+  title        = {tiny\_schiller: a small German Schiller corpus for tiny language models},
+  year         = {2023},
+  howpublished = {\url{https://github.com/schutera/tiny_schiller}},
+  note         = {Source texts: Projekt Gutenberg-DE, public domain.}
+}
+```
+
+If you use the underlying texts, please also acknowledge Projekt
+Gutenberg-DE (<https://www.projekt-gutenberg.org/>).
+
+## Limitations
+
+- No critical apparatus, no editorial notes, no philological annotations.
+- No part-of-speech, lemma, or syntactic annotations.
+- Eleven works only at v1; Schiller's poetry, ballads, philosophical
+  essays, and historical writing are not included.
+- Encoding artifacts from the source HTML may remain after cleaning.
+- The 90/10 split is a contiguous-tail split; not suitable as a
+  benchmark.
+- Not intended for production language modeling, modern-German
+  evaluation, or stylometric work.
+
+## References
+
+- Projekt Gutenberg-DE — Schiller author page:
+  <https://www.projekt-gutenberg.org/autoren/namen/schiller.html>
+- Projekt Gutenberg-DE — Impressum:
+  <https://www.projekt-gutenberg.org/info/impressum/impressum.html>
+- `tiny_shakespeare` (structural template):
+  <https://huggingface.co/datasets/tiny_shakespeare>
+- EU Directive 2006/116/EC (term of protection of copyright):
+  <https://eur-lex.europa.eu/eli/dir/2006/116/oj>

README.md

@@ -1,3 +1,132 @@
 ---
-license: cc0-1.0
+language:
+- de
+license:
+- cc0-1.0
+- mit
+license_details: "Text content (Schiller's works) is public domain (CC0). Pipeline code is MIT licensed."
+task_categories:
+- text-generation
+pretty_name: tiny_schiller
+size_categories:
+- 1M<n<10M
 ---
+
+# tiny_schiller
+
+Inspired by tiny_shakespeare, this brings you the works from the, as a matter of course, best German poet and writer - Friedrich Schiller, to your language models.
+
+*"Das Leben ist Nur ein Moment, der Tod ist auch nur einer!" - Friedrich Schiller*
+
+![Friedrich Schiller](schiller_thug.gif)
+
+## Description
+- Download Size: ~1,982 KB
+- Version 1.0.0 (no release notes)
+
+```python
+# Feature Structure
+FeaturesDict({
+    'text': Text(shape=(), dtype=string),
+})
+```
+
+55,629 lines or 327,233 words or 2,028,161 characters (including spaces) of Friedrich Schiller.
+The document comprises of the following works, originally from [Projekt Gutenberg](https://www.projekt-gutenberg.org/autoren/namen/schiller.html).
+
+
+## Quick Start
+
+Pick the tokenisation that fits your use case and run the corresponding prepare script — it downloads `tiny_schiller.clean.txt` automatically if needed and writes `train.bin` / `val.bin` to the same directory.
+
+```bash
+# Character-level (96-char vocab — smallest, best for teaching)
+python schiller_char/prepare.py
+
+# GPT-2 BPE (50k vocab, uint16 ids — compatible with nanoGPT defaults)
+python schiller_bpe/prepare.py
+
+# cl100k BPE (100k vocab, uint32 ids — 25 % fewer tokens on German text)
+python schiller_cl100k/prepare.py
+```
+
+Then point your nanoGPT config at the generated files, e.g.:
+
+```python
+# in your nanoGPT train.py or config
+dataset = 'schiller_char'   # or schiller_bpe / schiller_cl100k
+```
+
+See [`DATA_CARD.md`](DATA_CARD.md) for corpus statistics, tokenizer comparison, and licensing details.
+
+
+## HuggingFace Datasets
+
+```python
+from datasets import load_dataset
+
+# Full corpus as a single text field
+ds = load_dataset("schutera/tiny_schiller")
+
+# 90/10 train / test split (contiguous tail, same convention as tiny_shakespeare)
+ds = load_dataset("schutera/tiny_schiller", "train_test")
+```
+
+
+## Fine-tuning small LLMs
+
+The corpus works as a continued pre-training or domain adaptation target for any
+causal LM. An end-to-end example using TRL `SFTTrainer` is in
+[`examples/finetune_sft.py`](examples/finetune_sft.py):
+
+```bash
+pip install transformers trl datasets accelerate
+python examples/finetune_sft.py --model TinyLlama/TinyLlama-1.1B-Chat-v1.0
+```
+
+The default context window is **2048 tokens**. Pass `--context_length` to match your model's actual window — there is no config file, just one flag:
+
+```bash
+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
+```
+
+To pre-chunk the corpus to a JSONL file instead:
+
+```bash
+python scripts/chunk.py --size 2048                  # gpt2 tokenizer, default
+python scripts/chunk.py --size 4096 --tokenizer cl100k_base
+```
+
+Tested model targets: TinyLlama 1.1B, Phi-3 Mini 3.8B, Llama 3.2 1B/3B, Qwen2.5 0.5B–3B.
+
+
+## Dramen und begleitende Schriften:
+- Briefe über Don Carlos
+- Aus dem Szenar zum »Demetrius«
+- Demetrius
+- Die Huldigung der Künste
+- Die Jungfrau von Orleans
+- Die Räuber
+- Die Verschwörung des Fiesco zu Genua
+- Don Carlos, Infant von Spanien
+- Kabale und Liebe
+- Maria Stuart
+- Wallenstein
+- Wilhelm Tell
+
+## Citation
+
+In conclusion, the works of Friedrich Schiller are a valuable contribution to literature, and this project aims to make them more accessible to language models. We hope you find this resource useful for your research and creative endeavors.
+
+If you use this work in your research or projects, please cite it as follows:
+
+```markdown
+@misc{schutera2023tinyschiller,
+  author       = {Schutera, Mark},
+  title        = {tiny\_schiller: a small German Schiller corpus for tiny language models},
+  year         = {2023},
+  howpublished = {\url{https://github.com/schutera/tiny_schiller}},
+  note         = {Source texts: Projekt Gutenberg-DE, public domain.}
+}
+```

tiny_schiller.py

@@ -0,0 +1,117 @@
+# coding=utf-8
+# Copyright 2023 Mark Schutera. Licensed under MIT (pipeline code); text is public domain.
+#
+# Modelled after the tiny_shakespeare HuggingFace dataset loading script:
+# https://huggingface.co/datasets/tiny_shakespeare
+
+"""tiny_schiller — a small German Schiller corpus for tiny language models."""
+
+import datasets
+
+
+_DESCRIPTION = """\
+A clean, concatenated plain-text corpus of Friedrich Schiller's works, assembled
+for training tiny character-level and subword language models.  The source texts
+are in the public domain (Projekt Gutenberg-DE); the pipeline code is MIT-licensed.
+"""
+
+_HOMEPAGE = "https://github.com/schutera/tiny_schiller"
+
+_LICENSE = "Public domain (text) / MIT (pipeline code)"
+
+_CITATION = """\
+@misc{schutera2023tinyschiller,
+  author       = {Schutera, Mark},
+  title        = {tiny\\_schiller: a small German Schiller corpus for tiny language models},
+  year         = {2023},
+  howpublished = {\\url{https://github.com/schutera/tiny_schiller}},
+  note         = {Source texts: Projekt Gutenberg-DE, public domain.}
+}
+"""
+
+_DATA_URL = (
+    "https://huggingface.co/datasets/mrkschtr/tiny_schiller/resolve/main/tiny_schiller.clean.txt"
+)
+
+
+class TinySchillerConfig(datasets.BuilderConfig):
+    """BuilderConfig for tiny_schiller."""
+
+    def __init__(self, **kwargs):
+        super(TinySchillerConfig, self).__init__(
+            version=datasets.Version("1.0.0"),
+            **kwargs,
+        )
+
+
+class TinySchiller(datasets.GeneratorBasedBuilder):
+    """tiny_schiller: a small German Schiller corpus for tiny language models."""
+
+    BUILDER_CONFIGS = [
+        TinySchillerConfig(
+            name="default",
+            description=(
+                "Returns the entire corpus as a single example with a 'text' field "
+                "(train split only)."
+            ),
+        ),
+        TinySchillerConfig(
+            name="train_test",
+            description=(
+                "Splits the corpus 90 / 10 by character count (contiguous tail for test), "
+                "each split returned as a single example with a 'text' field.  "
+                "Follows the same convention as tiny_shakespeare."
+            ),
+        ),
+    ]
+
+    DEFAULT_CONFIG_NAME = "default"
+
+    def _info(self):
+        return datasets.DatasetInfo(
+            description=_DESCRIPTION,
+            features=datasets.Features({"text": datasets.Value("string")}),
+            supervised_keys=None,
+            homepage=_HOMEPAGE,
+            license=_LICENSE,
+            citation=_CITATION,
+        )
+
+    def _split_generators(self, dl_manager):
+        data_file = dl_manager.download_and_extract(_DATA_URL)
+
+        if self.config.name == "train_test":
+            return [
+                datasets.SplitGenerator(
+                    name=datasets.Split.TRAIN,
+                    gen_kwargs={"filepath": data_file, "split": "train"},
+                ),
+                datasets.SplitGenerator(
+                    name=datasets.Split.TEST,
+                    gen_kwargs={"filepath": data_file, "split": "test"},
+                ),
+            ]
+
+        # "default" — single train split containing the full corpus
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={"filepath": data_file, "split": "all"},
+            ),
+        ]
+
+    def _generate_examples(self, filepath, split):
+        with open(filepath, encoding="utf-8") as f:
+            corpus = f.read()
+
+        if split == "all":
+            yield 0, {"text": corpus}
+            return
+
+        # 90 / 10 contiguous-tail split (same convention as tiny_shakespeare)
+        split_index = int(len(corpus) * 0.9)
+
+        if split == "train":
+            yield 0, {"text": corpus[:split_index]}
+        else:  # split == "test"
+            yield 1, {"text": corpus[split_index:]}