Angad2005/Creative-Writing1

Creative Writing1

A chat-format instruction-tuning dataset focused mainly on creative writing, with smaller portions of code, translation, rewriting and reasoning tasks. Responses are long-form and include a reasoning section before the final answer.

Contents

Config	Split	File	Rows	Description
default	train	data/processed/train.jsonl	[N_PROCESSED]	Cleaned training set. The data the model was actually trained on (train_new.jsonl).
default	test	data/eval/evaluate.jsonl	[N_EVAL]	Held-out evaluation set.
original	train	data/original/train.jsonl	1,583	Original dataset before processing, kept for reproducibility.

.
├── README.md
└── data/
    ├── original/train.jsonl
    ├── processed/train.jsonl
    └── eval/evaluate.jsonl

Data format

Each line is one JSON object in chat format:

{
  "messages": [
    {"role": "system", "content": "You are an ..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "tools": []
}

Field	Type	Description
messages	list of {role, content}	The conversation. Roles are system, user, assistant.
tools	list	Tool definitions available to the conversation. [Describe, or state that it is unused / empty.]

Notes on structure (measured on the original file):

• 96.4% of examples include a system prompt.
• 99.9% are single-turn (one user message); one example has three user turns.
• Responses start with a reasoning section (a "think" block) followed by the final answer. [Specify the exact delimiter, e.g. <think>...</think>.]

Dataset statistics

Measured on the original train.jsonl with tiktoken (cl100k_base). [Re-run the audit on train_new.jsonl and evaluate.jsonl and update this section.]

Metric	Value
Usable records	1,583
Total tokens	6,312,571
Tokens per example (mean / median)	3,988 / 1,112
Tokens per example (p5 / p95 / max)	448 / 10,224 / 20,744
Prompt tokens (mean / median)	124.5 / 126
Response tokens (mean / median)	3,863 / 970
Examples over 4,096 tokens	647 (40.9%)

The length distribution is bimodal: roughly 59% of examples are under about 1,800 tokens, and most of the rest fall between 5,000 and 11,000 tokens. Set your maximum sequence length accordingly when training.

Task mix (keyword heuristic, approximate)

Task type	Share
Creative / writing	86.4%
Code	8.5%
Other	2.3%
Translation	1.6%
Math / reasoning	0.7%
Summarization	0.5%
Open QA	0.1%

All text is Latin script (English).

Usage

from datasets import load_dataset

# processed train + eval
ds = load_dataset("Angad2005/Creative-Writing1")
print(ds)
print(ds["train"][0]["messages"])

# original, unprocessed data
orig = load_dataset("Angad2005/Creative-Writing1", "original")

How it was built

1. Source: [Describe where the prompts and responses came from, e.g. generated with model X, written by hand, mixed.]
2. Processing (train.jsonl to train_new.jsonl): [Describe the steps: removing invalid JSON lines, deduplication, length filtering, removing refusals / boilerplate, truncation handling, etc.]
3. Evaluation set: [Describe how evaluate.jsonl was built and how it was kept separate from training data.]

Known limitations

An automated audit of the original train.jsonl found the following. Check whether each was fixed in the processed version and update this list.

• Heavily templated prompts. 94% of prompts start with "You are an...". Prompt diversity is low (distinct-2 of 0.37), so models may overfit to this phrasing.
• Skewed toward creative writing. About 86% of examples; do not expect strong coverage of code, math or factual QA.
• Repetitive response openings. Many reasoning sections begin with the same few phrases.
• Long examples. Over 40% exceed 4,096 tokens and will be truncated if you train with that limit.
• Data-quality issues in the original file: 8 invalid JSON lines, 40 exact duplicate copies (35 groups), 15 prompts that appear more than once with different responses, 22 refusal or AI-boilerplate responses, 20 possibly truncated responses, and 10 responses with highly repetitive n-grams.
• Contamination not verified against the evaluation set in the original audit. [Add result of the contamination check here.]

Intended use

Supervised fine-tuning of language models for long-form creative writing and reasoning-style responses. Not intended for use as a factual knowledge source.

License

[Specify the license and any restrictions that come from the source of the data or the model that generated it.]

Citation

@misc{angad2005_creativewriting1,
  title  = {Creative Writing1},
  author = {Angad2005},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Angad2005/Creative-Writing1}
}