Large Language Model Prompt Datasets: An In-depth Analysis and Insights

Yuanming Zhang, Yan Lin, Arijit Khan, Huaiyu WanSubmission to ICLR 2026 · Withdrawn

TL;DR

This paper provides the first comprehensive survey and taxonomy of LLM prompt datasets, curating 129 repositories comprising over 673 million instances and 1.22 TB of data. Through multi-level linguistic analysis (lexical, syntactic, and semantic), the authors demonstrate how prompt datasets fundamentally diverge from traditional text corpora. Building on these structural insights, they introduce a data-driven prompt optimization technique that steers prompt syntax toward a high-performing embedding centroid to improve LLM output accuracy.

What Problem Does This Paper Address?

While the machine learning community has heavily investigated pre-training and fine-tuning corpora, the instructions and queries that guide models at inference—prompts—have rarely been analyzed systematically. Prompt datasets are scattered across GitHub repositories, community forums (e.g., Reddit), and academic benchmarks with inconsistent schemas, varying generation techniques, and fragmented quality.

Consequently, practitioners lack a holistic understanding of how prompts are structured across different domains (e.g., general chat vs. business vs. clinical tasks) and how their grammatical properties differ from conventional prose like web text or literature. Furthermore, existing prompt engineering approaches are often heuristic, lacking a principled, data-driven way to refine the underlying syntax of a prompt.

Methods

  1. Systematic Curation & Taxonomy: The authors sourced datasets across platforms (Hugging Face, Kaggle, GitHub, forums), filtering down to 129 datasets via strict criteria (size thresholds $\ge 1\text{K}$ for standard or $\ge 50$ for user-shared, formatting cleanliness, clear task formulation, and research-friendly licenses). They built a hierarchical taxonomy covering Source (publisher, channel, generation process), Content (language, format, prompt engineering cues, extra metadata), and Target (cognitive intent, downstream applications).
  2. Multi-Level Linguistic Profiling:
  3. Lexical: Lemmatized $n$-gram modeling ($n=3,4,5$) and corpus-level TF-IDF vectorization with cosine similarity.
  4. Syntactic: Dependency parsing (using spaCy's en_core_web_sm), Part-of-Speech (POS) tagging, and verb–direct object (dobj) relation tracking.
  5. Semantic: Sentence-BERT dense embeddings (all-MiniLM-L6-v2) projected onto 2D space using Principal Component Analysis (PCA).
  6. Centroid-Based Prompt Optimization: High-dimensional embeddings of POS and dependency relations are computed across top-performing prompts to define a syntactic "centroid." For an underperforming target prompt, deviations from this centroid guide a structured modification plan, prompting the LLM to rewrite the query into an aligned syntactic form.

Datasets

The authors cataloged 129 prompt datasets, selecting seven representative English-only datasets for deep linguistic study:

  • 1100+ ChatGPT Prompts for Business: Curated business/marketing templates.
  • BoredHumans: Community-created creative and experimental prompts.
  • databricks-dolly-15k: Human-authored instruction-response pairs covering general tasks.
  • medical-o1-reasoning-SFT: Model-generated clinical reasoning questions distilled with GPT-4o.
  • OASST1: Crowd-sourced, tree-structured dialogue messages.
  • Self-Instruct: Synthetic prompt-response pairs iteratively generated from seed tasks.
  • ShareGPT: Real-world multi-turn conversation logs between users and ChatGPT.

Reference non-prompt corpora: Universal Dependencies English EWT (informal web/social) and ParTUT (formal legal/news/wiki), alongside Google Books $n$-grams.
Evaluation benchmarks for optimization: PRM800K and MMLU-Pro.

Key Insights and Ideas

  • Inquiry vs. Exposition: Prompts fundamentally diverge from standard corpora. While literary corpora (Google Books) feature descriptive connectors (e.g., "at the end of the"), prompt datasets are dominated by imperative action verbs and direct framing (e.g., "what be the", "I want you to").
  • Domain Dictates Syntax:
  • Medical prompts prioritize descriptive specificity: high adjectival modifiers (amod = 0.107) and numerals (NUM = 0.031) paired with linking verbs to describe states, lesions, and symptoms.
  • Business prompts favor concise, actionable directives: high direct objects (dobj = 0.090) and low determiners (det = 0.051), clustering around imperative pairs like "create plan" and "increase traffic".
  • Dataset Artifacts: Repetitive synthetic templates introduce severe distributional skews. Self-Instruct showed repetitive artifact $n$-grams from test sentences ("the quick brown fox jumps..."), while Dolly-15k showed disproportionate clustering around generic verb-noun stems like "tell i" and "give list".
  • Semantic Dispersion: Model-generated synthetic datasets (Self-Instruct) span a wide, uniform semantic footprint, whereas human-generated sets (Dolly, OASST1, ShareGPT) cluster closely together with significant overlap in PCA space.

Key Results

  • Linguistic Distribution: POS tagging showed nouns account for 25.9%–26.4% of domain-specific datasets (medical-o1, 1.1k-business), exceeding formal reference corpora like ParTUT (21.0%). Verb frequency reached up to 13.0% in directive business sets.
  • TF-IDF Vocabulary Overlap: OASST1 and ShareGPT exhibited the strongest lexical cosine similarity (0.82), driven by shared organic conversational phrasing, whereas Self-Instruct remained lexically distant from domain-specific sets (0.08 similarity to medical-o1).
  • Optimization Case Studies: In benchmark tasks on PRM800K and MMLU-Pro, syntactically aligning prompts with the POS/dependency centroid turned incorrect initial answers into correct reasoning steps on gpt-4o-mini (e.g., solving a circular permutation seating problem from an incorrect $5{,}040$ to the correct $10{,}080$).

Limitations and Future Work

  • Linguistic Scope: In-depth grammatical analysis was limited to seven English-language datasets, omitting non-English and cross-modal prompts present in the broader 129-dataset collection.
  • Empirical Scale of Prompt Optimization: The syntactic centroid rewriting approach was demonstrated via specific qualitative case studies rather than exhaustive benchmark-wide scoring across multiple LLM families.
  • Future Directions: The authors suggest extending this syntactic filtering framework to automatically detect malformed or off-distribution prompts prior to inference, as well as applying it to quality assessment and pricing models in emerging AI prompt marketplaces.

Summary updated

Abstract

A prompt is a natural language instruction that defines a specific task for a large language model (LLM) and serves as the primary interface for human-LLM interaction. With the growing deployment of LLMs, diverse prompt datasets are emerging from platforms such as GitHub and social media. These datasets span a wide array of applications and content types, facilitating both broader LLM utilization and improved prompt engineering. In this work, we--for the first time--have compiled an extensive list of prompt datasets sourced from various channels, representing a spectrum of downstream tasks, languages, engineering techniques, attributes, and modalities. We select key representative datasets for systematic analysis, revealing commonalities and differences in prompt construction across categories, distinguishing them from other text corpora like literature and web. We further propose a prompt optimization approach that leverages syntactic embeddings of part-of-speech and dependency structures. By identifying a centroid representation of prompts and guiding LLMs to rewrite prompts toward this centroid, our method improves the meaningfulness of model outputs. We have made our datasets and code available at \url{https://anonymous.4open.science/r/LLM-Prompt-Datasets-7416}.
BibTeX
@misc{zhang2025large,
  author = {Yuanming Zhang and Yan Lin and Arijit Khan and Huaiyu Wan},
  title = {{Large Language Model Prompt Datasets: An In-depth Analysis and Insights}},
  year = {2025},
  eprint = {2510.09316},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.09316},
}
Figure