DATA_CARD.md

# DATA_CARD — tiny_schiller (agent-ready)

This file is a compact, copy/paste-friendly summary of the **tiny_schiller** corpus and how to load it.

## What this is

- **Corpus**: 11 public-domain dramatic works by Friedrich Schiller (German)
- **Goal**: a *drop-in* German analogue to Karpathy’s *tiny_shakespeare* for small-LM prototyping, fine-tuning, and education
- **Release artifact**: a cleaned, single UTF-8 text file + deterministic preprocessing scripts

## Quantitative summary

**Corpus file (cleaned)**

- File: `tiny_schiller.parsed.txt`
- Characters: 2,019,857
- Bytes (UTF-8): 2,067,041 (~2.07 MB)
- Encoding: UTF-8 (no BOM)
- Line endings: LF
- Unique codepoints: 88
- Guillemets `«` / `»`: 101 / 101
- Works: 11

**Precomputed token streams (90/10 train/val)**

- `schiller_char/`
  - Vocab: 88
  - Train tokens: 1,817,871
  - Val tokens: 201,986
  - chars/token: 1.00
- `schiller_bpe/` (GPT-2 BPE)
  - Vocab: 50,257
  - Train tokens: 768,768
  - Val tokens: 85,844
  - Total tokens: 854,611
  - chars/token: 2.36
- `schiller_cl100k/` (`cl100k_base`)
  - Vocab: ~100k
  - Train tokens: 578,182
  - Val tokens: 64,412
  - Total tokens: 642,593
  - chars/token: 3.14

**Fine-tuning parquet files (Hub: `data/`)**

- Whole-work split: `data/train.parquet`, `data/test.parquet`
  - Rows: 9 train works, 2 test works
  - Fields: `title`, `text`
- Instruction-formatted dialogue completion: `data/instruct.parquet`
  - Rows: 7,607
  - Fields: `prompt`, `completion`, `work`, `character`
- Per-character persona splits: `data/char_*.parquet`
  - Files: 89
  - Example row counts: `char_WALLENSTEIN.parquet` 264, `char_CARLOS.parquet` 235, `char_MOOR.parquet` 130

**Reference fine-tune (paper)**

- Model: `Qwen/Qwen2.5-0.5B-Instruct` (494M params)
- Hardware: NVIDIA RTX 3060 (12 GB VRAM), bf16
- Stage 1: 2 epochs on `instruct.parquet`, batch 4, context 512, lr 2e-4
- Stage 2: 2 epochs on `char_MOOR.parquet`, lr 1e-4, weight decay 0.05
- Wall-clock time: ~2.7 hours total (~2.6 h stage 1, <3 min stage 2)
- Reported outcome: 99.05% token accuracy, entropy 0.055 (two epochs, weight decay 0.05)
- Recipe: `examples/reference_finetune.py` (two-stage run)

## Canonical text file

- `tiny_schiller.parsed.txt`
  - UTF-8, no BOM
  - LF line endings
  - Canonical speaker turns in the form `SPEAKER:\ntext`

## Precomputed tokenization splits

These match common small-LM workflows (e.g. nanoGPT-style `train.bin`/`val.bin`).

- `schiller_char/` — character-level (88 unique characters)
- `schiller_bpe/` — GPT-2 BPE via `tiktoken`
- `schiller_cl100k/` — `cl100k_base` via `tiktoken`

Build locally:

```bash
python schiller_char/prepare.py
python schiller_bpe/prepare.py
python schiller_cl100k/prepare.py
```

## HuggingFace datasets (whole works)

Dataset id: `mrkschtr/tiny_schiller`

Each row is one work with at least `title` and `text` fields.

```python
from datasets import load_dataset

ds = load_dataset("mrkschtr/tiny_schiller")
print(ds["train"][0]["title"])
print(ds["train"][0]["text"][:200])
```

Split policy (paper/README): train holds out **Wilhelm Tell** and **Die Braut von Messina** as test.

## Instruction + per-character persona splits (parquet)

These ship on the Hub under `data/`.

- General dialogue-completion: `data/instruct.parquet` (7,607 rows)
- Per-character persona files: `data/char_*.parquet` (89 files)

Schema per row:

- `prompt`: instruction template + preceding context turns
- `completion`: next speaker turn in `NAME:\ntext` format
- `work`: work title
- `character`: normalized speaker id (uppercase, underscores)

Load examples:

```python
from datasets import load_dataset

# Instruction-formatted dialogue completions
ins = load_dataset(
    "mrkschtr/tiny_schiller",
    data_files="data/instruct.parquet",
    split="train",
)

# One character persona dataset
moor = load_dataset(
    "mrkschtr/tiny_schiller",
    data_files="data/char_MOOR.parquet",
    split="train",
)
```

Rebuild locally (writes to `data/` by default):

```bash
python scripts/build_instruct.py
python scripts/build_instruct.py --list-characters
python scripts/build_instruct.py --character MOOR
```

## Intended use

- Rapid prototyping of model/training/tokenization choices under tight compute
- Stylistic fine-tuning on a homogeneous German literary register
- Per-character persona fine-tuning using the prebuilt `char_*.parquet` splits
- Education and reproducibility baselines on a non-English “tiny corpus”

## Licensing (read this)

- **Text content** (`tiny_schiller.txt`, `tiny_schiller.parsed.txt`): public domain (Schiller died 1805) and redistributed from DraCor / GerDraCor under CC0
- **Code + documentation** (everything else, incl. scripts and this file): MIT

See `LICENSING.md` for details and links to the upstream CC0 claim.