Synthetic-archive / README.md
croqaz's picture
Update README.md
5f9efc5 verified
|
Raw
History Blame Contribute Delete
6.43 kB
metadata
pretty_name: The Synthetic Archive
language:
  - en
tags:
  - synthetic
  - vintage
  - historical
  - text-generation
  - question-answering
  - text-rewriting
  - language-modeling
task_categories:
  - text-generation
  - question-answering
  - summarization

The Synthetic Archive

Work in progress

Generation is still WIP. The final number of records, compressed size, and token count will be reported when the release is complete.

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 manageable 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 is to create varied language-model training material from 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..."}

The generation program:

  1. Reads valid records from seeds.jsonl.
  2. Splits each seed into chunks of approximately 3,000 tokens.
  3. Uses at most three chunks per seed.
  4. Applies each generation template to each chunk.
  5. Sends the requests to a vLLM server serving Gemma-4-E4B.
  6. Appends successful generations to a JSONL output file.

Generation tasks

The production command used five prompt templates.

continue

Produces a clean continuation of the source passage. The model is instructed not to summarize or comment on the source and to continue it as prose.

diverse QA pairs

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.

extract knowledge

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.

narrative

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

lost manuscript

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

Generation model

The generations were produced with Gemma-4-E4B.

The generation settings were:

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

Limitations

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

Generated text may contain errors

Generated records may:

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

Question-and-answer pairs have not necessarily been checked by a human or independently verified against the source.

Historical authenticity

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

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.

The seed texts may contain:

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