croqaz/Synthetic-archive

The Synthetic Archive

Synthetic Archive is a large synthetic English-text dataset generated from OCR-derived historical and period-style passages. Knowledge cutoff is year 1900.

Each source passage was divided into chunks and processed through several generation tasks, including:

• generating continuations of unfinished passages;
• creating question-and-answer pairs;
• extracting and reformulating factual knowledge;
• rewriting material as a narrative;
• transforming source material into a period-style magazine feature;
• reconstructing or continuing an apparently incomplete historical manuscript.

The goal was to create varied materials from original archival text while retaining useful information, literary variety, historical vocabulary, and different forms of textual reasoning.

Inspiration

This project is inspired by the synthetic-data generation methodology described in the Fine-Phrase paper and by the use of synthetic data in the training of small language models, including SmolLM3-3B. This is an independent project. It is not an official Fine-Phrase, Hugging Face, or SmolLM3 dataset.

Dataset composition

The source material is supplied as JSON Lines. Each valid seed record has the form:

{"text": "A historical or period-style passage...",
  "extra": {"prompt": "magazine_feature", "seed_index": 13130, "chunk_index": 0, "source_preview": "..."}}

The generation steps:

1. Read valid records from seeds.jsonl.
2. Split each seed into chunks of ~3,000 tokens.
3. Use at most three chunks per seed.
4. Apply each generation template to each chunk.
5. Send the requests to a vLLM server serving Gemma-4-E4B.
6. Append successful generations to a JSONL output file.

Generation tasks

The pipeline used five prompt templates (included here for reference).

lost manuscript prompt

Generates a continuation in the style of an older manuscript, preserving period diction, syntax, tone, and historical plausibility.

extract knowledge prompt

Rewrites factual knowledge from the source into clear, educational prose. The prompt asks the model to preserve the source's information and supporting explanations without adding new facts.

magazine prompt

Transforms the document into a magazine-style feature article with an engaging opening, with flowing prose.

narrative prompt

Reorganizes the source into a clear narrative emphasizing temporal order, processes, causes, consequences, and relationships between events.

diverse QA pairs prompt **

Produces up to eight question-and-answer pairs based on the source. The prompt requests several question types, including factual, open-ended, yes/no, comparative, multiple-choice, comprehension, and problem-solving questions where appropriate.

** This is currently broken and not included here; if I manage to fix it I'll upload it.

Generation model

The generations were produced with Gemma-4-E4B. Generation settings were:

┌─────────────────────────┬─────────────────┐
│ Setting                 │ Value           │
├─────────────────────────┼─────────────────┤
│ Inference mode          │ Chat completion │
├─────────────────────────┼─────────────────┤
│ Temperature             │ 0.8             │
├─────────────────────────┼─────────────────┤
│ Top-p                   │ 0.95            │
├─────────────────────────┼─────────────────┤
│ Maximum output          │ 1,024 tokens    │
├─────────────────────────┼─────────────────┤
│ Chunk budget            │ ~ 3,000 tokens  │
├─────────────────────────┼─────────────────┤
│ Maximum chunks per seed │ 3               │
└─────────────────────────┴─────────────────┘

Stats

Metric	Mean	Min	P5	P95	Max
Char length	2,318.4	55.0	944.0	4,906.0	8,717.0
Unique chars	39.03	8.00	30.00	48.00	63.00
Token count	375.3	6.0	152.0	805.0	1,024.0
Sentences	15.85	1.00	6.00	36.00	147.00
Quality	104.72	50.00	100.91	107.38	113.73
Compression	98.81	51.48	93.37	103.60	149.06
Entropy	98.32	75.08	95.78	101.42	121.18
Dict hit	97.63	10.53	94.61	99.63	100.00
Alpha	99.47	50.48	98.41	100.00	100.00
Vowel	99.74	50.00	98.47	100.00	100.00
ASCII	99.78	50.00	98.84	100.00	100.00
Score	96.74	-133.07	89.92	101.20	107.21

Limitations

This is a synthetic dataset and should not be treated as a verified reference corpus.

The generated text is historical-style / historically influenced text, not authenticated historical writing. It may combine period vocabulary with some modern reasoning, modern assumptions, or factual mistakes.

Generated text may contain errors

Generated records may:

• hallucinate information;
• misinterpret OCR errors;
• produce unsupported answers;
• distort the meaning of a source;
• imitate historical style inaccurately;
• end abruptly due to the output-token limit;
• contain modern assumptions in otherwise historical-sounding prose.

Source material

The seeds are generated from the highest quality texts filtered from a super-dataset containing croqaz/vintage-v1, croqaz/vintage-v2, haykgrigorian/english-historical-corpus-1800-1875 and jbduran/think-dataset-clean.

seeds.jsonl file contains 4,550,770 records, English-only, pre-year-1900 (included here for reference).

The seed texts usually contain:

• line-break hyphenation;
• incorrectly recognized characters;
• broken words;
• unusual punctuation;
• stray OCR marks;
• inconsistent capitalization;
• historical spelling and vocabulary;
• outdated terminology and attitudes.

Citation

If you find this dataset valuable, please consider citing:

@misc{synthetic-archive,
  title  = {The Synthetic Archive},
  author = {Cristi Constantin},
  month  = {June},
  year   = {2026},
  url = {https://huggingface.co/datasets/croqaz/Synthetic-archive}
}