Skip to main content
The Deltarune Complete Transcript Dataset is a structured collection of dialogue and narration extracted from Deltarune Chapters 1–4. It covers all major scenes and characters and is provided in three machine-readable formats suitable for data analysis, search, and language model fine-tuning. The dataset contains between 10,000 and 100,000 records across all chapters.

​
File coverage

FileCoverageStatus
data/chap1_cleaned.txt / chap1_dataset.jsonlChapter 1 (full)Stable
data/chap2_cleaned.txt / chap2_dataset.jsonlChapter 2 (full, Normal Route)Stable
data/chap3_cleaned.txt / chap3_dataset.jsonlChapter 3 (full, includes Sword Route)Stable
data/chap4_cleaned.txt / chap4_dataset.jsonlChapter 4 (full, Normal Route)Beta
Chapter 4 data is currently in Beta. Record counts and scene ordering may change in future releases.

​
Available formats

The dataset is distributed in three formats. Each serves a different use case.
FormatLocationBest for
JSONLdata/chap*_dataset.jsonlStreaming, scripting, custom pipelines
Parquetparquet/chap*_dataset.parquet, parquet/full_chapters_dataset.parquetAnalytical queries, columnar access, large-scale filtering
ChatMLdata/chatml/deltarune_story_chatml.jsonlInstruction fine-tuning of language models

JSONL format

Newline-delimited JSON records. One dialogue line per record with context, speaker, and text fields.

Parquet format

Columnar binary format for fast filtering and aggregation. Includes a combined file covering all chapters.

ChatML format

Instruction-tuning format with system, user, and assistant roles. Use for fine-tuning language models.

​
Repository structure

data/
  chap1_cleaned.txt         # Cleaned text, Chapter 1
  chap1_dataset.jsonl       # Structured JSONL, Chapter 1
  chap2_cleaned.txt
  chap2_dataset.jsonl
  chap3_cleaned.txt
  chap3_dataset.jsonl
  chap4_cleaned.txt
  chap4_dataset.jsonl
  chatml/
    deltarune_story_chatml.jsonl  # ChatML format, all chapters

parquet/
  chap1_dataset.parquet
  chap2_dataset.parquet
  chap3_dataset.parquet
  chap4_dataset.parquet
  full_chapters_dataset.parquet  # All chapters combined

raw/
  chap1.txt   # Raw pre-processed transcript
  chap2.txt
  chap3.txt
  chap4.txt

parquet.py    # Script to regenerate parquets from JSONL

​
The raw/ directory

The raw/ directory contains the original pre-processed transcript text files (chap1.txt through chap4.txt). These are unstructured plain text files and are not suitable for direct programmatic use. They represent the source material before scene segmentation, speaker attribution, and field extraction were applied. Use the data/chap*_cleaned.txt files if you need a human-readable cleaned version, or any of the structured formats above for programmatic access.