satyapratheek/dostoyevsky_chunks

Dostoyevsky Chunks Dataset

This dataset contains preprocessed text chunks from four major works by Fyodor Dostoyevsky, prepared for fine-tuning language models on the author's distinctive writing style.

Dataset Description

Dataset Summary

A curated collection of text segments extracted from public domain English translations of Dostoyevsky's novels, processed into consistent chunks suitable for causal language modeling tasks.

Supported Tasks

• Causal Language Modeling: Primary use case for training autoregressive language models
• Author Style Transfer: Learning Dostoyevsky's literary style and narrative techniques
• Text Generation: Generating text in the style of Dostoyevsky

Languages

• English: All texts are English translations of original Russian works

Dataset Structure

Data Instances

Each instance contains a single text chunk:

{
  "chunks": "But what can be expected of a man who has such a desperate character that he will hang himself for being always late for dinner..."
}

Data Fields

• chunks (string): A preprocessed text segment of approximately 512 tokens

Data Splits

The dataset contains 6,217 text chunks in a single split, suitable for:

• Training: Use 90% for fine-tuning
• Validation: Reserve 10% for evaluation

Source Data

Initial Data Collection

The source texts were downloaded from Project Gutenberg, ensuring all content is in the public domain.

Source Works

1. Crime and Punishment (Project Gutenberg #2554)
2. The Brothers Karamazov (Project Gutenberg #28054)
3. The Idiot
4. Notes from the Underground (Project Gutenberg #600)

Data Processing Pipeline

1. Header/Footer Removal: Used gutenberg-cleaner to remove Project Gutenberg boilerplate
2. Text Normalization: Applied ftfy for encoding fixes and unidecode for ASCII conversion
3. Content Filtering:
   • Removed paragraphs shorter than 200 characters
   • Filtered out chapter titles and section headers
   • Excluded epigraphs and other metadata
4. Chunking: Segmented text into approximately 512-token chunks using SmolLM tokenizer
5. Quality Control: Manual review of sample chunks for content quality

Dataset Creation

Curation Rationale

This dataset was created to enable fine-tuning of small language models on Dostoyevsky's writing style, characterized by:

• Psychological depth and introspection
• Philosophical themes and existential questions
• Complex character development
• Rich narrative voice and distinctive prose rhythm

Source Data Analysis

• Total chunks: 6,217
• Average chunk length: ~512 tokens
• Content diversity: Narrative prose, dialogue, internal monologue, philosophical passages
• Temporal coverage: Spans Dostoyevsky's major creative periods

Considerations for Using the Data

Social Impact of Dataset

This dataset preserves and makes accessible the literary heritage of one of world literature's greatest authors, enabling:

• Educational applications: Literary analysis and style studies
• Creative writing assistance: Learning narrative techniques
• Cultural preservation: Maintaining access to classic literature

Discussion of Biases

• Historical context: Reflects 19th-century social norms and perspectives
• Translation artifacts: Based on English translations, may not capture original Russian nuances
• Selection bias: Limited to four major works, may not represent complete stylistic evolution
• Cultural specificity: Reflects Russian cultural context and Orthodox Christian themes

Other Known Limitations

• Temporal scope: Limited to works from 1864-1880
• Genre limitation: Primarily novels, lacks short stories and journalism
• Translation dependency: Quality dependent on translator choices
• Processing artifacts: Some text segmentation may split related passages

Additional Information

Dataset Curators

• Created by: satyapratheek
• Processing date: July 2025
• Contact: Available through Hugging Face profile

Licensing Information

All source texts are in the public domain (published before 1928). The dataset processing and compilation is released under public domain as well.

Citation Information

@dataset{dostoyevsky_chunks_2025,
  title={Dostoyevsky Chunks: Preprocessed Text Dataset for Author Style Fine-tuning},
  author={satyapratheek},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/satyapratheek/dostoyevsky_chunks}
}

Contributions

If you use this dataset, please consider citing both the dataset and the original Project Gutenberg sources.

Usage Example

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("satyapratheek/dostoyevsky_chunks")

# Access the chunks
chunks = dataset["train"]["chunks"]
print(f"Total chunks: {len(chunks)}")
print(f"Sample chunk: {chunks[:200]}...")

# Use for fine-tuning
```python
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM-135M")

def tokenize_function(examples):
    return tokenizer(examples["chunks"], truncation=True, padding="max_length", max_length=512)

tokenized_dataset = dataset.map(tokenize_function, batched=True)

Related Models

This dataset was used to fine-tune:

• satyapratheek/smollm-dostoyevsky - LoRA fine-tuned SmolLM-135M

---

Processing Stats from Fine-tuning:

• Training duration: 4 hours, 56 minutes, 35 seconds
• Final training loss: 3.254
• Training samples/second: 1.048
• Hardware: Apple M1 MacBook Air (8GB RAM)