Skip to main content
The ChatML file packages the entire dataset as multi-turn conversations structured for instruction fine-tuning. Each record corresponds to one scene and contains three messages: a system prompt defining the assistant’s role, a user request for a specific scene transcript, and an assistant response with the full scene dialogue. This is the recommended format for fine-tuning language models on this dataset. The file is located at data/chatml/deltarune_story_chatml.jsonl.

​
Message structure

Each record is a JSON object with a single messages array containing three entries. 
{
  "messages": [
    {
      "role": "system",
      "content": "You are the Deltarune Script Archive. You provide exact, chronological scene transcripts including dialogues, actions, and pauses."
    },
    {
      "role": "user",
      "content": "Provide the transcript for Scene: Obj Krisroom.\nContext: This scene occurs after 'Device Contact' and before 'Obj Carcutscene'."
    },
    {
      "role": "assistant",
      "content": "Toriel: \"Kris...!\"\nToriel: \"Wake up!\"\nToriel: \"KRIS...!\"\nToriel: \"Kris, if you do not wake up, we will be late for school!\"\nToriel: \"I will wait outside for you, alright?\""
    }
  ]
}

​
Role descriptions

​
system

The system message establishes the assistant’s identity as the Deltarune Script Archive. It instructs the model to provide exact, chronological scene transcripts including dialogue, actions, and pauses. This prompt is identical across all records.

​
user

The user message requests a specific scene transcript. It follows this template: 
Provide the transcript for <scene name>.
Context: This scene occurs after '<previous scene>' and before '<next scene>'.
The context line gives the model positional information about where the scene falls in the story’s chronology.

​
assistant

The assistant message contains the full scene transcript. Formatting rules:
  • Each line of dialogue is formatted as Speaker: "line text"
  • One line per message, separated by newlines
  • Player choices are written as > [Player Choice: X]
  • Narration and actions follow the same Speaker: "text" pattern with Narrator as the speaker

​
Loading the data

import json

records = []
with open("data/chatml/deltarune_story_chatml.jsonl", "r", encoding="utf-8") as f:
    for line in f:
        record = json.loads(line)
        records.append(record)

print(f"Total records: {len(records)}")

# Inspect the first record
first = records[0]
for message in first["messages"]:
    print(f"--- {message['role']} ---")
    print(message["content"])
    print()

​
Inspect system and user prompts

import json

with open("data/chatml/deltarune_story_chatml.jsonl", "r", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

# Extract all user prompts (scene requests)
for record in records[:5]:
    messages = {m["role"]: m["content"] for m in record["messages"]}
    print(messages["user"])
    print()

​
Convert to HuggingFace datasets format

from datasets import Dataset
import json

# Load raw records
with open("data/chatml/deltarune_story_chatml.jsonl", "r", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

# Flatten into a list of dicts with 'messages' key
# HuggingFace datasets expects a dict of lists
data = {"messages": [record["messages"] for record in records]}

ds = Dataset.from_dict(data)
print(ds)
print(ds[0])

# Optionally push to HuggingFace Hub
# ds.push_to_hub("your-username/deltarune-chatml")
The messages field structure is compatible with the format expected by libraries such as trl (SFTTrainer), axolotl, and LLaMA-Factory when using their ChatML or conversation templates.

​
Use for fine-tuning

When fine-tuning, apply the model’s chat template to format the messages array before tokenization. Most fine-tuning frameworks handle this automatically when you pass a dataset with a messages column.
Example using trl and transformers: 
from datasets import Dataset
from trl import SFTTrainer, SFTConfig
from transformers import AutoTokenizer, AutoModelForCausalLM
import json

# Load dataset
with open("data/chatml/deltarune_story_chatml.jsonl", "r", encoding="utf-8") as f:
    records = [json.loads(line) for line in f]

ds = Dataset.from_dict({"messages": [r["messages"] for r in records]})

# Load model and tokenizer
model_id = "your-base-model-id"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

# Train
trainer = SFTTrainer(
    model=model,
    args=SFTConfig(output_dir="output"),
    train_dataset=ds,
    processing_class=tokenizer,
)
trainer.train()