dpaleka/WildChat-2k-TypeTopic

WildChat-2k-TypeTopic: The Manually Curated Edition

Dataset Description

WildChat-2k-TypeTopic is a manually curated subset of 1,880 real-world user prompts from the WildChat dataset, featuring annotations for both task type (e.g. knowledge recall, problem solving, creative, lists) and topic category (e.g. personal assistance, math, ai, household)

Why this dataset?

Suppose you want to answer a research question such as "What kind of user prompt does the LLM like doing most?" or "What is the implicit utility function of the LLM for answering different user prompts?" or "What kind of user prompts do models bail on"? The first step is to find a dataset of user prompts.

WildChat-1M is the most frequently used dataset of user prompts to LLMs; unfortunately everyone who has ever looked into it knows it is full of nonsensical prompts, typos, non-English, NSFW stuff, and other noise; and that the distribution of prompts that users ask for is very detailed in some domains (e.g. creative writing) and very sparse in others.

WildChat-2k-TypeTopic is a curated subset of single-message user prompts, constructed as follows:

1. 1.Filter out (using an LLM filter) prompts that:
2. 2.are not in English
3. 3.are not meaningful tasks (e.g., random character strings, “hello”)
4. 4.are incomplete (e.g., “Fix this code” with no code provided)
5. 5.are infeasible for text-only LLMs (e.g., “Describe a time when you worked in a team”, “an image of a cat”)
6. 6.are clearly part of multi-turn conversations (e.g., text-based game setups)
7. 7.are more than 800 characters long

1. 1.Deduplicate using text-embedding-3-large embeddings.
2. 2.Classify into 16 task types and 25 topic categories, then subsample ~2000 tasks to preserve representation of all types and categories.
3. 3.Manual review to remove anything problematic according to the described criteria.

Key Features

• —1,880 annotated prompts from real user interactions
• —15 task type categories (e.g., creative, coding, explanation, problem_solving)
• —24 topic categories (e.g., programmingother, creativewriting, personal_assistance)
• —Short prompts: 12-800 characters (median: 116)
• —Quality filtered: All entries are coherent English prompts, as opposed to WildChat

Dataset Structure

Data Format

The dataset is provided in JSONL format (newline-delimited JSON), with each entry containing:

{
  "id": "wildchat2k_0003",
  "text": "I want to learn how to understand and speak spanish, can you use the pareto principle, which identifies 20% of the topic that will yield 80% of the desired results, to create a learning plan for me?",
  "type": "planning_design",
  "topic": "languages"
}

Fields

• —id (string): Unique identifier
• —text (string): The user prompt/query
• —type (string): Task classification (15 categories)
• —topic (string): Subject matter classification (24 categories)

Task Types (15 categories)

Topic Categories (24 categories)

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("dpaleka/wildchat-2k-typetopic")

# Access the data
for item in dataset['train']:
    print(f"Type: {item['type']}, Topic: {item['topic']}")
    print(f"Text: {item['text']}\n")

Citation

If you use this dataset, please cite the original WildChat paper:

@inproceedings{zhao2024wildchat,
  title={WildChat: 1M ChatGPT Interaction Logs in the Wild},
  author={Zhao, Wenting and Havaldar, Savvas and Besmens, Harshita and Chiu, Ting-Hao 'Kenneth' and Pyatkin, Valentina and Lin, Bill Yuchen and Yu, Liwei and Liu, Alane Suhr and Zhang, Yejin and others},
  booktitle={ICLR},
  year={2024}
}

For this specific annotated subset:

@dataset{wildchat2k_typetopic,
  title={WildChat-2k-TypeTopic: Curated Subset of Single-Message User Prompts},
  author={Paleka, Daniel},
  year={2025},
  publisher={HuggingFace},
  url={https://huggingface.co/datasets/dpaleka/wildchat-2k-typetopic}
}