pith. sign in

arxiv: 2510.09316 · v2 · submitted 2025-10-10 · 💻 cs.LG · cs.CL

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

Yuanming Zhang , Yan Lin , Arijit Khan , Huaiyu Wan This is my paper

Pith reviewed 2026-05-18 08:25 UTC · model grok-4.3

classification 💻 cs.LG cs.CL
keywords LLM prompt datasetssyntactic featuresprompt routinglinguistic analysisprompt filteringdomain classificationprompt quality
0
0 comments X

The pith

Syntactic features recover over 93 percent of GPU embedding accuracy for routing LLM prompts while cutting latency by nearly half with no hardware required.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper compiles 129 heterogeneous LLM prompt datasets into a taxonomy and analyzes linguistic patterns across seven representative corpora at lexical, syntactic, and semantic levels. It establishes that 62-dimensional syntactic features based on part-of-speech and dependency distributions function as an efficient routing primitive. These features support downstream tasks such as prompt filtering, domain classification, and quality prediction without invoking any additional model. A sympathetic reader would care because the approach enables lightweight, low-latency prompt management in production systems that would otherwise rely on expensive GPU embeddings.

Core claim

The paper's central claim is that 62-d syntactic features consisting of POS tag and dependency parse distributions serve as a uniquely efficient routing primitive. These features recover more than 93 percent of the accuracy obtained by GPU-based embedding methods for prompt routing tasks, while delivering 1.9 times lower single-request latency at 3.0 ms versus 5.7 ms, all without any GPU or corpus vocabulary.

What carries the argument

62-dimensional syntactic feature vectors drawn from part-of-speech tag distributions and dependency parse statistics, deployed as a lightweight routing primitive.

Load-bearing premise

The seven representative corpora capture the linguistic properties of the full collection of 129 datasets and the observed feature patterns generalize to unseen prompt distributions.

What would settle it

Measure routing accuracy of the 62-d syntactic features against GPU embeddings on a new collection of LLM prompt datasets drawn from sources outside the original 129 and check whether accuracy remains above 93 percent.

Figures

Figures reproduced from arXiv: 2510.09316 by Arijit Khan, Huaiyu Wan, Yan Lin, Yuanming Zhang.

Figure 1
Figure 1. Figure 1: The hierarchical taxonomy of prompt datasets a single method (e.g., awesome-chatgpt-prompts uses role playing), while others combine multiple techniques (e.g., PromptBench (Zhu et al., 2024) integrates six methods), or leave the strategy unspecified. Moreover, datasets may include extra attributes, including labeled data (e.g., response supervision, safety labels), analytical data (e.g., token count, user … view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of 3/4/5-grams in the same dataset and 5-grams across multiple datasets. The ratio is defined as the count of the specific n-gram divided by the count of prompts in the dataset. More comprehensive comparison data and analysis can be found in Appendix F.1. Analysis of results. The n-gram frequency distributions reveal several notable patterns that highlight the distinct functional and stylistic c… view at source ↗
Figure 3
Figure 3. Figure 3: (a-b): The top-10 most common verbs and their top-5 direct noun objects in two prompt datasets. Data for other 5 datasets are shown in [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Semantic prompt embeddings distribution. Analysis of results. (1) Wide coverage in Self￾Instruct: The Self-Instruct dataset exhibits the most dispersed and evenly distributed semantic space, suggesting a broad topical coverage. This aligns with the self-instruction paradigm’s goal of generating diverse instruction types. (2) Se￾mantic cohesion in specific domains: Prompts from medical-o1 and 1.1k-business … view at source ↗
Figure 5
Figure 5. Figure 5: A case study of prompt optimization To illustrate the practical impact of this approach, we present one representative case study in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: The conclusions drawn from these figures are consistent with the main paper, for example, high￾frequency n-grams phrases mainly show command sentences and topic content. In addition, there are two other findings: 1. The n-grams phrases of some datasets include abnormal content (e.g. “identify which instrument be string” in dolly-15 and “The quick brown fox jumps over the lazy dog” in Self-Instruct), which … view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of 3/4/5-grams in the same dataset 45 [PITH_FULL_IMAGE:figures/full_fig_p045_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Top-5 n-grams comparison across datasets F.2 SYNTACTIC-LEVEL ANALYSIS In this section, we present the complete experimental data for all identified dependency types, along with their proportions in the datasets, as shown in [PITH_FULL_IMAGE:figures/full_fig_p046_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The top-10 most common verbs and their top-5 direct noun objects in prompt datasets. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Semantic prompt embeddings distribution for all other datasets. 49 [PITH_FULL_IMAGE:figures/full_fig_p049_9.png] view at source ↗
read the original abstract

We compile 129 heterogeneous LLM prompt datasets (>1.22 TB, >673M instances) into a structured taxonomy and conduct a multi-level linguistic analysis (lexical, syntactic, and semantic) on seven representative corpora, surfacing systematic patterns that distinguish prompts from general text. Three downstream experiments validate practical utility: prompt filtering (F1 = 0.90), domain classification (Macro-F1 = 0.975), and prompt quality prediction (AUC = 0.792), all without invoking any additional model. A central finding is that 62-d syntactic features (POS + dependency distributions) serve as a uniquely efficient routing primitive, recovering >93% of GPU-embedding accuracy at 1.9 $\times$ lower single-request latency (3.0 ms vs. 5.7 ms) with no GPU and no corpus vocabulary. A complementary discriminative--predictive divergence shows that features most useful for routing are precisely those most negatively correlated with response quality, while lexical diversity (Cohen's $d$ = 0.71) dominates the quality signal but carries minimal routing weight, directly motivating two-stage pipeline design. Our datasets and code are available.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 1 minor

Summary. The paper compiles 129 heterogeneous LLM prompt datasets (>1.22 TB, >673M instances) into a structured taxonomy and conducts a multi-level linguistic analysis (lexical, syntactic, and semantic) on seven representative corpora. It validates practical utility through three downstream experiments using standard linguistic tools: prompt filtering (F1 = 0.90), domain classification (Macro-F1 = 0.975), and prompt quality prediction (AUC = 0.792), all without invoking additional models. A central finding is that 62-d syntactic features (POS + dependency distributions) recover >93% of GPU-embedding accuracy for routing at 1.9× lower single-request latency (3.0 ms vs. 5.7 ms) with no GPU and no corpus vocabulary. The work also reports a discriminative-predictive divergence, where features useful for routing are negatively correlated with response quality while lexical diversity (Cohen's d = 0.71) dominates the quality signal, motivating two-stage pipeline designs. Datasets and code are made available.

Significance. If the results hold, the manuscript offers a large-scale empirical characterization of LLM prompts and demonstrates efficient, model-free methods for prompt routing, filtering, and classification using compact syntactic features. Concrete metrics (F1=0.90, Macro-F1=0.975, AUC=0.792, >93% recovery) from off-the-shelf parsers, together with code availability, support verification. The discriminative-predictive divergence provides a concrete rationale for separating routing and quality stages in LLM serving systems. These contributions could inform prompt dataset curation and low-latency routing primitives in production environments.

major comments (1)
  1. The generalization of the 62-d syntactic feature performance (recovering >93% of embedding accuracy) and the observed patterns from the seven representative corpora to the full collection of 129 datasets and to unseen prompts is load-bearing for the broad utility and routing-primitive claims (abstract). The manuscript should supply explicit selection criteria for the seven corpora, quantitative comparisons of their distributional properties (e.g., length, domain, lexical diversity) against the full >1.22 TB collection, and any statistical controls or hold-out validation demonstrating representativeness. Absent this, the reported latency advantage and discriminative-predictive divergence may not transfer reliably.
minor comments (1)
  1. Abstract: the term 'discriminative--predictive divergence' is introduced without a brief parenthetical definition; a short clarification on first use would aid readers.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and the recommendation for minor revision. The concern about representativeness of the seven corpora is well-taken and directly relevant to the strength of our generalization claims. We address it point-by-point below and will incorporate the requested details in the revision.

read point-by-point responses

  1. Referee: The generalization of the 62-d syntactic feature performance (recovering >93% of embedding accuracy) and the observed patterns from the seven representative corpora to the full collection of 129 datasets and to unseen prompts is load-bearing for the broad utility and routing-primitive claims (abstract). The manuscript should supply explicit selection criteria for the seven corpora, quantitative comparisons of their distributional properties (e.g., length, domain, lexical diversity) against the full >1.22 TB collection, and any statistical controls or hold-out validation demonstrating representativeness. Absent this, the reported latency advantage and discriminative-predictive divergence may not transfer reliably.

    Authors: We agree that explicit documentation of selection criteria and distributional comparisons is necessary to support the generalization claims. The seven corpora were chosen to maximize coverage across the taxonomy categories (instruction, dialogue, reasoning, creative, domain-specific) while balancing computational feasibility for full linguistic parsing; they include both large-scale public datasets and smaller curated ones to span the observed range of prompt lengths and sources. In the revision we will add: (1) a dedicated subsection stating the selection criteria with a table listing each corpus, its size, primary domain, and taxonomy category; (2) quantitative comparisons (mean/median token length, type-token ratio as lexical diversity proxy, domain label distribution, and syntactic feature variance) between the seven and summary statistics computed over the full 129 datasets using available metadata; (3) a hold-out experiment in which syntactic-feature classifiers trained on the seven are evaluated on a random 10% sample drawn from the remaining 122 datasets, reporting F1 and latency metrics to quantify transfer. These additions will be placed in Section 4 and the appendix. We believe the requested material can be supplied without altering the core results. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical measurements on external datasets

full rationale

The paper compiles 129 external prompt datasets and performs direct multi-level linguistic analysis on seven representative corpora using off-the-shelf POS and dependency parsers. All reported metrics (prompt filtering F1=0.90, domain classification Macro-F1=0.975, quality prediction AUC=0.792, >93% recovery of GPU-embedding accuracy, 3.0 ms vs 5.7 ms latency) are obtained from straightforward feature extraction and downstream task evaluation on the collected data. No equations, fitted parameters, or self-citations reduce these results to inputs defined within the paper itself. The generalization assumption from seven corpora to the full collection is a validity concern rather than a circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Work rests on standard NLP preprocessing tools and the assumption that the sampled corpora represent broader prompt distributions; no new entities or ad-hoc parameters are introduced beyond the derived 62-dimensional feature vector.

axioms (1)
  • domain assumption POS tagging and dependency parsing tools produce reliable distributions on prompt text.
    Invoked when constructing the 62-d syntactic feature vectors used for routing and quality prediction.

pith-pipeline@v0.9.0 · 5740 in / 1214 out tokens · 33738 ms · 2026-05-18T08:25:39.329438+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

138 extracted references · 138 canonical work pages

  1. [1]

    1100+ ChatGPT Prompts for Business

    1100+ ChatGPT Prompts for Business • Publisher: Chris Porter 15 • Size: 1235 instances • License: - • Link: https://chatgpt-business-prompts.notion.site/ 1100-ChatGPT-Prompts-for-Business-eea03b0bc9b84ae7a5bdbd76a67460f3 • Description: "1100+ ChatGPT Prompts for Business" is a Notion-based dataset containing 1,235 curated prompts tailored for diverse busi...

    work page

  2. [2]

    Each entry typically includes a prompt, an automatic prompt (system prompt like), and a concise description

    2.5k-chatgpt-promp-templates • Publisher: TheVeller • Size: 1088 instances • License: - • Link: https://ignacio-velasquez.notion.site/ 2-500-ChatGPT-Prompt-Templates-d9541e901b2b4e8f800e819bdc0256da • Description: This dataset comprises over 1,000 curated ChatGPT prompt templates in Notion Workspace format, spanning diverse domains such as AI, marketing, ...

    work page

  3. [3]

    A Collection of AI’s Prompts for optimal context • Publisher: Marc-Aurele Besner • Size: 70 instances • License: MIT • Link: https://github.com/marc-aurele-besner/ ChatGPT-PromptsList • Description: This repository offers a well-curated collection of conversation prompts tailored for OpenAI’s GPT-3 model

    work page

  4. [4]

    Academic Reasoning and Intuition Chains Dataset • Publisher: Marco De Santis • Size: 2024 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/marcodsn/academic-chains • Description: The Academic Reasoning and Intuition Chains dataset comprises 1,975 ex- amples of chain-of-thought reasoning distilled from open-access arXiv papers across...

    work page 2024

  5. [5]

    Each prompt is available in multiple languages, enabling cross-linguistic studies of prompt effectiveness and translation consistency

    AI Short • Publisher: rockbenben • Size: 5867 instances • License: - • Link: https://www.aishort.top/ • Description: AI Short is a public prompt-sharing platform with 5,867 categorized prompts. Each prompt is available in multiple languages, enabling cross-linguistic studies of prompt effectiveness and translation consistency

    work page

  6. [6]

    Stored as JSON objects, it enables research in synthetic prompt generation, model creativity evaluation, and downstream fine-tuning

    AI-Generated Prompts Dataset • Publisher: Anthony Therrien 16 • Size: 173574 instances • License: CC-BY-SA-4.0 • Link: https://www.kaggle.com/datasets/anthonytherrien/ ai-generated-prompts-dataset • Description: This dataset features thousands of prompts generated by the teknium/OpenHermes-2p5-Mistral-7B model, each designed to elicit diverse and contextu...

    work page

  7. [7]

    Its user-driven structure offers valuable insights into real-world prompt usage, preferences, and task design patterns

    AIPRM • Publisher: AIPRM • Size: 5325 instances • License: - • Link: https://www.aiprm.com/ • Description: AIPRM is a community-curated prompt library and management platform featuring 5,325 publicly accessible prompts categorized by topic and activity. Its user-driven structure offers valuable insights into real-world prompt usage, preferences, and task ...

    work page

  8. [8]

    Designed for fine-tuning LLaMA models, it enables research in alignment, instruction tuning, and synthetic data generation

    Alpaca_data • Publisher: Stanford Alpaca • Size: 52K instances • License: Apache-2.0 • Link: https://github.com/tatsu-lab/stanford_alpaca/tree/main • Description: The Stanford Alpaca dataset comprises 52K high-quality, instruction-following examples generated via a modified Self-Instruct pipeline using text-davinci-003. Designed for fine-tuning LLaMA mode...

    work page

  9. [9]

    It comprises 48,818 examples, each featuring an instruction, optional input context, and a GPT-4-generated response, facilitating text- generation and fine-tuning tasks

    Alpaca_GPT4_data_zh • Publisher: Microsoft Research • Size: 52K instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/llm-wizard/ alpaca-gpt4-data-zh • Description: Alpaca_GPT4_data_zh is a Chinese instruction-tuning dataset curated by the Instruction Tuning with GPT-4 project. It comprises 48,818 examples, each featuring an instruction,...

    work page

  10. [10]

    Each query includes four samples from three models (1.5B, 7B, and R1), with pass rates computed per model to assign unbi- ased difficulty scores

    AM-DeepSeek-Distilled-40M • Publisher: a-m-team • Size: 40M instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/a-m-team/ AM-DeepSeek-Distilled-40M • Description: AM-DeepSeek-Distilled-40M is a multilingual (zh/en) reasoning dataset com- prising 3.34 million prompts paired with 40 million model-generated responses across code, math, ...

    work page

  11. [11]

    Collected from diverse open-source sources, it features semantically deduplicated instructions spanning text, code, and math domains

    AM-DeepSeek-R1-Distilled-1.4M 17 • Publisher: a-m-team • Size: 1.4M instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/a-m-team/ AM-DeepSeek-R1-Distilled-1.4M • Description: AM-DeepSeek-R1-Distilled-1.4M is a bilingual (Chinese and English) rea- soning dataset of 1.4 million challenging problem-solution pairs. Collected from diverse...

    work page

  12. [12]

    It contains 100k+ problems from repositories and categorized by pass rates of Qwen models

    AM-Math-Difficulty-RL • Publisher: a-m-team • Size: 234729 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/a-m-team/ AM-Math-Difficulty-RL • Description: AM-Math-Difficulty-RL is an English math dataset comprising three difficulty tiers designed for RL of LLMs. It contains 100k+ problems from repositories and categorized by pass ...

    work page

  13. [13]

    APIGen-MT -5k • Publisher: Salesforce AI Research • Size: 5K instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/Salesforce/APIGen-MT-5k • Description: The APIGen-MT-5k dataset comprises 5000 realistic, high-quality, multi-turn function-calling dialogues generated by APIGen-MT, a scalable automated agentic pipeline simulating agent-h...

    work page

  14. [14]

    Featuring both human- and LLM-generated entries with clear attribution, it supports research in prompt engineering, prompt effectiveness, and cross-model generalization

    awesome-chatgpt-prompts • Publisher: Fatih Kadir Akın • Size: 211 instances • License: CC0-1.0 • Link: https://github.com/f/awesome-chatgpt-prompts • Description: The Awesome ChatGPT Prompts dataset is a collaboratively curated collection of diverse prompts optimized for interactive AI models, including ChatGPT, Claude, and LLaMA. Featuring both human- an...

    work page

  15. [15]

    Aya Collection • Publisher: Cohere For AI Community et al. • Size: 513M instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/CohereLabs/aya_ collection 18 • Description: Aya Collection is a massive multilingual instruction tuning dataset comprising over 513 million prompt-completion pairs across 115 languages. It integrates three source...

    work page

  16. [16]

    Aya Dataset • Publisher: Cohere For AI Community et al. • Size: 204K instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/CohereLabs/aya_dataset • Description: The Aya Dataset is a multilingual, human-annotated instruction fine-tuning resource encompassing 204K prompt-completion pairs across 65 languages and dialects. It includes origin...

    work page

  17. [17]

    needle-in-a-haystack

    BABILong • Publisher: AIRI et al. • Size: 25K instances • License: Apache 2.0 • Link: https://huggingface.co/datasets/RMT-team/babilong • Description: BABILong is a generative benchmark designed to evaluate large language mod- els’ ability to perform reasoning over extremely long contexts. It embeds the ten bAbI tasks within irrelevant PG19 background tex...

    work page

  18. [18]

    It builds upon 67K unique English prompts drawn from Alpaca and Dolly, automatically translated via Google Translate into 51 languages

    Bactrain-X • Publisher: MBZUAI • Size: 3484884 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/MBZUAI/Bactrian-X • Description: Bactrian-X is a multilingual instruction-following dataset containing 3.4 million instruction-input-response triplets across 52 languages. It builds upon 67K unique English prompts drawn from Alpaca and ...

    work page

  19. [19]

    Baize • Publisher: University of California et al. • Size: 210311 instances • License: GPL-3.0 • Link: https://huggingface.co/datasets/linkanjarad/ baize-chat-data • Description: Baize Chat Data is an instruction-finetuning corpus combining four sources: Alpaca, Medical, Quora, and StackOverflow. It contains about 210,000 conversational exam- ples, each f...

    work page

  20. [20]

    400k personalized Chinese character dialogues generated by the BELLE project

    BELLE_Generated_Chat • Publisher: BELLE • Size: 396004 instances • License: GPL-3.0 • Link: https://huggingface.co/datasets/BelleGroup/generated_ chat_0.4M • Description: BELLE_Generated_Chat contains approx. 400k personalized Chinese character dialogues generated by the BELLE project. Each record includes an instruction, an (empty) input, and a generated...

    work page

  21. [21]

    Hu- man:

    BELLE_Multiturn_Chat • Publisher: BELLE • Size: 831036 instances • License: GPL-3.0 • Link: https://huggingface.co/datasets/BelleGroup/multiturn_ chat_0.8M • Description: BELLE_Multiturn_Chat is a Chinese multi-turn conversational dataset com- prising approximately 0.8 million human-assistant dialogues generated by the BELLE project using ChatGPT. Each re...

    work page

  22. [22]

    It includes human-assistant exchanges across 13 instruction categories

    BELLE_train_3.5M_CN • Publisher: BELLE • Size: 3606402 instances • License: GPL-3.0 • Link: https://huggingface.co/datasets/BelleGroup/train_3.5M_CN • Description: The BELLE_train_3.5M_CN dataset comprises approximately 3.5 million monolingual Chinese instruction-response pairs generated by the BELLE project, format- ted as multi-turn and single-turn dial...

    work page

  23. [23]

    It provides real multi-model response comparisons (e.g., GPT-4, ChatGPT, NewBing, Wenxin) and continuous updates via collaborative platforms

    best-chinese-prompt • Publisher: K-Render • Size: 141 instances • License: - • Link: https://github.com/K-Render/best-chinese-prompt • Description: The Best Chinese Prompt dataset is a comprehensive, well-structured collection of Chinese-language prompts spanning diverse categories such as casual chat, knowledge Q&A, creative planning, copywriting, and co...

    work page

  24. [24]

    BigDocs-Bench • Publisher: ServiceNow Research et al. • Size: 415740 instances 20 • License: CC-BY-4.0 • Link: https://huggingface.co/datasets/ServiceNow/BigDocs-Bench • Description: BigDocs-Bench is a CC-BY-4.0 benchmark suite for training and evaluating multimodal models on document and code tasks. It comprises seven configurations: GUI- VQA, GUI2BBox, ...

    work page

  25. [25]

    BoredHumans • Publisher: Impulse Communications, Inc. • Size: 964 instances • License: - • Link: https://boredhumans.com/prompts.php • Description: BoredHumans is a diverse and extensive prompt dataset compiled from multiple sources, including Awesome ChatGPT Prompts, Data Science Prompts, and Tree-of-Thought Prompting, among others. Its rich variety cove...

    work page

  26. [26]

    CAMEL • Publisher: KAUST • Size: 1659328 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/camel-ai/ai_society • Description: CAMEL AI Society is a synthetic dialogue corpus comprising 25,000 simulated conversations between GPT-3.5-turbo agents role-playing across 50 distinct user roles and 50 assistant roles on ten tasks per pairi...

    work page

  27. [27]

    It enables research on prompt engineering techniques, model behavior across different AI platforms, and strategies for enhancing response quality

    ChatGPT & Bing AI Prompts • Publisher: yokoffing • Size: 35 instances • License: CC0-1.0 • Link: https://github.com/yokoffing/ChatGPT-Prompts • Description: The ChatGPT & Bing AI Prompts dataset offers a diverse collection of prompts designed to optimize interaction with advanced conversational AI models, including ChatGPT and Bing AI. It enables research...

    work page

  28. [28]

    It facilitates research on natural language interfaces for data analysis, model explanation, and automation of complex workflows

    ChatGPT Data Science Prompts • Publisher: Travis Tang • Size: 60 instances • License: - • Link: https://github.com/travistangvh/ChatGPT-Data-Science-Prompts • Description: The ChatGPT Prompts for Data Science dataset offers a curated collection of specialized prompts designed to enhance AI applications in data science tasks. It facilitates research on nat...

    work page

  29. [29]

    ChatGPT Prompts 21 • Publisher: PrathamKumar14 • Size: 84 instances • License: - • Link: https://github.com/PrathamKumar14/ChatGPT-Prompts • Description: The ChatGPT-Prompts dataset compiles diverse prompt templates focused on educational and productivity applications, including tutoring in web development, algorithm explanation, Excel formulas, social me...

    work page

  30. [30]

    Its value lies in providing versatile, real-world prompt examples that support research on prompt engineering and AI interaction across various domains

    ChatGPT Prompts • Publisher: ColorblindAdam • Size: 19 instances • License: - • Link: https://github.com/ColorblindAdam/ChatGPTPrompts • Description: The ChatGPT Prompts dataset offers a broad collection of prompts covering diverse topics, designed for use with GPT 3.5. Its value lies in providing versatile, real-world prompt examples that support researc...

    work page

  31. [31]

    These prompts serve multiple research purposes, including natural language generation, prompt engineering, and AI-driven creativity

    ChatGPT Prompts • Publisher: Matheus Nunes Puppe • Size: 36 instances • License: - • Link: https://github.com/puppe1990/useful_chatgpt_prompts/ blob/main/src/promptsData.js • Description: The ChatGPT Prompts dataset originates from a web application offering a diverse set of prompts generated by OpenAI’s GPT-3 model. These prompts serve multiple research ...

    work page

  32. [32]

    Chinese-DeepSeek-R1-Distill-data-110k • Publisher: Cong Liu et al. • Size: 110K instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/Congliu/ Chinese-DeepSeek-R1-Distill-data-110k • Description: Chinese-DeepSeek-R1-Distill-data-110k is a 110K-entry Chinese dataset dis- tilled from DeepSeek-R1, supporting text generation, text2text gener...

    work page

  33. [33]

    Chinese-DeepSeek-R1-Distill-data-110k-SFT • Publisher: Cong Liu et al. • Size: 110K instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/Congliu/ Chinese-DeepSeek-R1-Distill-data-110k-SFT • Description: Licensed under Apache-2.0, Chinese-DeepSeek-R1-Distill-data-110k-SFT is an open-source, Chinese-language instruction-tuning dataset dis...

    work page

  34. [34]

    original

    CoCoNot • Publisher: Allen Institute for AI et al. • Size: 13784 instances • License: ODC-BY-1.0 • Link: https://huggingface.co/datasets/allenai/coconot • Description: CoCoNot is a novel English dataset for benchmarking and improving contextual noncompliance in chat-based language models. It offers three configurations: “original” contains 11K training an...

    work page

  35. [35]

    COIG-CQIA • Publisher: Shenzhen Institute of Advanced Technology et al. • Size: 44694 instances • License: - • Link: https://huggingface.co/datasets/m-a-p/COIG-CQIA • Description: COIG-CQIA (Chinese Open Instruction Generalist - Quality is All You Need) is a high-quality, open-source Chinese instruction tuning dataset designed to align language models wit...

    work page

  36. [36]

    Each sample includes a locally posed query, its English translation, four answer options in both languages, and metadata such as image source, license, category, and a unique ID

    CVQA • Publisher: MBZUAI • Size: 10374 instances • License: Mixed • Link: https://huggingface.co/datasets/afaji/cvqa • Description: CVQA is a culturally diverse, multilingual visual question-answering bench- mark featuring over 10,000 image-based questions across 39 country-language pairs. Each sample includes a locally posed query, its English translatio...

    work page

  37. [37]

    Provided under a CC-BY-SA 3.0 license, this English-language dataset supports academic or commercial use

    databricks-dolly-15K • Publisher: Databricks • Size: 15011 instances • License: CC-BY-SA-3.0 • Link: https://huggingface.co/datasets/databricks/ databricks-dolly-15k • Description: Databricks-dolly-15K is an open-source corpus of over 15,000 human- generated instruction-response pairs created by Databricks employees across eight behavioral categories defi...

    work page

  38. [38]

    DeepMath-103K 23 • Publisher: Tencent et al. • Size: 103110 instances • License: MIT • Link: https://huggingface.co/datasets/zwhe99/DeepMath-103K • Description: DeepMath-103K is a large-scale, MIT-licensed dataset comprising 103K chal- lenging mathematical problems tailored for text-to-text and text-generation tasks. Each example includes a problem statem...

    work page

  39. [39]

    It comprises 8 million formal statements and corresponding proofs generated from high-school and undergraduate-level mathematical contest problems

    DeepSeek-Prover-V1 • Publisher: DeepSeek • Size: 27503 instances • License: deepseek-license • Link: https://huggingface.co/datasets/deepseek-ai/ DeepSeek-Prover-V1 • Description: DeepSeek-Prover-V1 is a large-scale synthetic proof dataset for Lean 4 theo- rem proving. It comprises 8 million formal statements and corresponding proofs generated from high-s...

    work page

  40. [40]

    DialogStudio • Publisher: Salesforce AI et al. • Size: 87 datasets • License: Apache-2.0 • Link: https://huggingface.co/datasets/Salesforce/dialogstudio • Description: DialogStudio is a large-scale, unified collection of dialogue datasets curated to advance conversational AI. It integrates a wide range of domains—such as task-oriented dialogue, open-domai...

    work page

  41. [41]

    DMind_Benchmark • Publisher: Zhejiang Univerisity et al. • Size: 1869 instances • License: - • Link: https://huggingface.co/datasets/DMindAI/DMind_Benchmark • Description: DMind_Benchmark is a comprehensive dataset for evaluating large language models on blockchain, cryptocurrency, and Web3 knowledge. It provides objective (mul- tiple choice) and subjecti...

    work page

  42. [42]

    Dynosaur • Publisher: UCLA et al. 24 • Size: 801900 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/Dynosaur/ dynosaur-sub-superni • Description: Dynosaur introduces a dynamic and low-cost paradigm for curating instruction- tuning datasets. It automatically generates diverse instructions by leveraging metadata from HuggingFace data...

    work page

  43. [43]

    Exploring the Possibilities of AI Prompts Over 200 Ideas

    Exploring the Possibilities of AI Prompts Over 200 Ideas • Publisher: Muhammad Bilal • Size: 165 instances • License: MIT • Link: https://github.com/bilalnawaz072/AI-Prompts-200-Ideas • Description: "Exploring the Possibilities of AI Prompts Over 200 Ideas" is a comprehen- sive dataset featuring over 200 prompts spanning diverse marketing and content crea...

    work page

  44. [44]

    1M • Description: Firefly is a Chinese instruction-tuning dataset comprising 1.15 million high- quality examples drawn from 23 common Chinese natural language processing datasets

    Firefly • Publisher: YeungNLP • Size: 1649399 instances • License: - • Link: https://huggingface.co/datasets/YeungNLP/firefly-train-1. 1M • Description: Firefly is a Chinese instruction-tuning dataset comprising 1.15 million high- quality examples drawn from 23 common Chinese natural language processing datasets. Each example includes a task type, an inpu...

    work page

  45. [45]

    Originating with FLAN 2021 and expanded in the FLAN Collection, this resource supports research on fine-tuning methods that enable large models to better follow human instructions

    Flan 2021 • Publisher: Google Research • Size: 62 datasets • License: Apache-2.0 • Link: https://github.com/google-research/FLAN • Description: The FLAN Instruction Tuning Repository provides datasets and code to generate instruction tuning collections that improve language model generalization and zero-shot performance. Originating with FLAN 2021 and exp...

    work page 2021

  46. [46]

    Each task is provided in zero-/few-shot and option/no-option formats as JSONL entries including inputs, targets, and task identifiers

    Flan 2022 • Publisher: Google Research • Size: 1836 datasets • License: Apache-2.0 • Link: https://huggingface.co/datasets/SirNeural/flan_v2 25 • Description: This dataset aggregates tasks from Flan, T0, Super-Natural Instructions, Chain- of-Thought, and Dialog into a training split. Each task is provided in zero-/few-shot and option/no-option formats as ...

    work page 2022

  47. [47]

    Flan-mini • Publisher: Singapore University of Technology and Design • Size: 1.34M instances • License: CC • Link: https://huggingface.co/datasets/declare-lab/flan-mini • Description: Flan-mini is a curated 1.34 M-example subset of the FLAN instruction-tuning collection augmented with code and conversational tasks. It pools 388K Flan2021 in- structions, 3...

    work page

  48. [48]

    Developed alongside the Step1X-Edit framework, it emphasizes real-world usage scenarios and supports a diverse array of image- to-image editing tasks

    GEdit-Bench • Publisher: StepFun • Size: 1212 instances • License: MIT • Link: https://huggingface.co/datasets/stepfun-ai/GEdit-Bench • Description: GEdit-Bench is a novel benchmark dataset designed to facilitate authentic evaluation of general-purpose image editing models. Developed alongside the Step1X-Edit framework, it emphasizes real-world usage scen...

    work page

  49. [49]

    It pairs user prompts with AI-generated replies and source metadata, covering various topics and styles

    GPT4All • Publisher: nomic-ai • Size: 739259 instances • License: MIT • Link: https://huggingface.co/datasets/nomic-ai/gpt4all_prompt_ generations • Description: The GPT4All dataset comprises 437,604 English prompt-response pairs drawn from diverse sources to facilitate training and fine-tuning of open-source text generation mod- els. It pairs user prompt...

    work page

  50. [50]

    GraphWalks • Publisher: OpenAI • Size: 1150 instances • License: MIT • Link: https://huggingface.co/datasets/openai/graphwalks • Description: GraphWalks is an open-source benchmark dataset designed to evaluate multi- hop reasoning over long graph contexts. Released under the MIT license, it provides directed graphs as edge lists alongside user-specified o...

    work page

  51. [51]

    It contains a main configuration and a Socratic variant, each offering questions and answers with calculator annotations and step-by-step reasoning expressed in natural language

    GSM8K • Publisher: OpenAI • Size: 17584 instances • License: MIT • Link: https://huggingface.co/datasets/openai/gsm8k • Description: GSM8K (Grade School Math 8K) is an English monolingual dataset of 8.8K crowd-sourced grade school math word problems paired with multi-step solutions. It contains a main configuration and a Socratic variant, each offering qu...

    work page

  52. [52]

    HARDMath • Publisher: Harvard University • Size: 1060 instances • License: MIT • Link: https://github.com/sarahmart/HARDMath • Description: HARDMath is a benchmark dataset designed to evaluate advanced mathe- matical reasoning in large language models, focusing on challenging graduate-level applied mathematics problems. Unlike existing benchmarks that emp...

    work page

  53. [53]

    HC3 • Publisher: SimpleAI • Size: 37175 instances • License: CC-BY-SA-4.0 • Link: https://huggingface.co/datasets/Hello-SimpleAI/HC3 • Description: The Human ChatGPT Comparison Corpus (HC3) is the first large-scale bilin- gual dataset enabling direct comparison of human and ChatGPT-generated text. Spanning English and Chinese samples, it encompasses betwe...

    work page

  54. [54]

    It includes paired comparison data from base and iterated models, as well as red teaming transcripts designed to expose model vulnerabilities

    hh-rlhf • Publisher: Anthropic • Size: 14M instances • License: MIT • Link: https://github.com/anthropics/hh-rlhf • Description: hh-rlhf provides valuable human preference data focused on helpfulness and harmlessness for training safer AI assistants using Reinforcement Learning from Human Feedback. It includes paired comparison data from base and iterated...

    work page

  55. [55]

    InstructDial • Publisher: Carnegie Mellon University • Size: 59 datasets • License: Apache-2.0 • Link: https://github.com/prakharguptaz/Instructdial 27 • Description: InstructDial is a comprehensive instruction tuning framework designed to improve zero-shot and few-shot generalization in dialogue systems. It unifies 48 diverse dialogue tasks from 59 datas...

    work page

  56. [56]

    Unlike previous synthetic datasets, InstructWild emphasizes authentic, varied user intents without relying on self-generated instructions

    InstructionWild_v1 • Publisher: National University of Singapore • Size: 104K instances • License: Non-Commercial Research Purpose • Link: https://github.com/XueFuzhao/InstructionWild • Description: InstructWild is a large-scale, user-sourced instruction dataset comprising over 110K high-quality, diverse instructions collected from real ChatGPT usage shar...

    work page

  57. [57]

    InstructionWild_v2 • Publisher: National University of Singapore • Size: 110K instances • License: Non-Commercial Research Purpose • Link: https://github.com/XueFuzhao/InstructionWild

    work page

  58. [58]

    Intellect-2-RL-Dataset • Publisher: PrimeIntellect • Size: 284741 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/PrimeIntellect/ INTELLECT-2-RL-Dataset • Description: Intellect-2-RL-Dataset is a large-scale collection of 284,741 training examples, designed for reinforcement learning in mathematical and coding problem solving. Each...

    work page

  59. [59]

    LaMini-instruction • Publisher: Monash University et al. • Size: 2585615 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/MBZUAI/ LaMini-instruction • Description: LaMini-Instruction is an English text-to-text generation dataset comprising 2.58M instruction-response pairs distilled from GPT-3.5-Turbo. Each sample includes an instr...

    work page

  60. [60]

    LCCC • Publisher: Tsinghua University et al. • Size: 12M instances • License: MIT • Link: https://huggingface.co/datasets/thu-coai/lccc 28 • Description: LCCC (Large-scale Cleaned Chinese Conversation Corpus) is a monolingual Chinese dialogue dataset with over 12 million conversations collected from social media. A strict and rigorous cleaning pipeline—in...

    work page

  61. [61]

    LIMA-sft • Publisher: Meta AI et al. • Size: 1330 instances • License: CC-BY-NC-SA • Link: https://huggingface.co/datasets/GAIR/lima • Description: The LIMA dataset contains 1,000 high-quality prompt-response pairs designed to align language models with the style of a helpful AI assistant. Prompts are diverse, sourced from Stack Exchange, wikiHow, Writing...

    work page

  62. [62]

    It includes over 33M SFT examples across code, math, science, chat, and safety, plus 56K instruction-following RL examples

    Llama-Nemotron-Post-Training-Dataset • Publisher: NVIDIA • Size: 33011757 instances • License: CC-BY-4.0 • Link: https://huggingface.co/datasets/nvidia/ Llama-Nemotron-Post-Training-Dataset • Description: The Llama-Nemotron-Post-Training-Dataset is a comprehensive dataset of synthetic SFT and RL samples designed to bolster reasoning, code, math, science, ...

    work page

  63. [63]

    LMSYS-Chat-1M • Publisher: UC Berkeley et al. • Size: 1M instances • License: LMSYS-Chat-1M license • Link: https://huggingface.co/datasets/lmsys/lmsys-chat-1m • Description: LMSYS-Chat-1M is a large-scale dataset of one million real-world LLM conver- sations, collected from 210K users interacting with 25 models via Chatbot Arena and Vicuna demo (April-Au...

    work page 2023

  64. [64]

    LongForm • Publisher: LMU Munich et al. • Size: 27739 instances • License: MIT • Link: https://huggingface.co/datasets/akoksal/LongForm • Description: LongForm is a 27K-example English instruction-following dataset under MIT license, for tasks like table QA, summarization, text generation, question answering. It collects human-written documents from C4 (1...

    work page

  65. [65]

    Spanning 21 categories from arithmetic to topology and logic, it offers human-verified, step-by-step reasoning examples in parallel languages

    Math_CoT_Arabic_English_Reasoning • Publisher: Miscovery AI • Size: 2834 instances • License: MIT • Link: https://huggingface.co/datasets/miscovery/Math_CoT_ Arabic_English_Reasoning • Description: Math CoT Arabic English Reasoning is a bilingual dataset of 1K-10K meticu- lously curated English and Arabic math problems with explicit chain-of-thought solut...

    work page

  66. [66]

    medical-o1-reasoning-SFT • Publisher: The Chinese University of Hong Kong, Shenzhen et al. • Size: 90120 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/FreedomIntelligence/ medical-o1-reasoning-SFT • Description: medical-o1-reasoning-SFT is a supervised fine-tuning dataset designed to enhance advanced medical reasoning in HuatuoGP...

    work page

  67. [67]

    medical-o1-verifiable-problem • Publisher: The Chinese University of Hong Kong, Shenzhen et al. • Size: 40644 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/FreedomIntelligence/ medical-o1-verifiable-problem • Description: medical-o1-verifiable-problem is an Apache-2.0 licensed dataset comprising open-ended medical reasoning probl...

    work page

  68. [68]

    Medical-R1-Distill-Data • Publisher: The Chinese University of Hong Kong, Shenzhen et al. • Size: 22000 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/FreedomIntelligence/ Medical-R1-Distill-Data • Description: Medical-R1-Distill-Data is an Apache-2.0 licensed instruction fine-tuning dataset distilled from Deepseek-R1’s Full Power...

    work page

  69. [69]

    thinking paths

    MedReason • Publisher: UC Santa Cruz et al. • Size: 32682 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/UCSC-VLAA/MedReason • Description: MedReason is a large-scale medical reasoning dataset combining seven clinical question-answer sources with a structured knowledge graph to produce detailed chains of reasoning. It contains 32,...

    work page

  70. [70]

    Medtrinity-25M • Publisher: Huazhong University of Science and Technology et al. • Size: 24922190 instances • License: Mixed • Link: https://huggingface.co/datasets/UCSC-VLAA/MedTrinity-25M • Description: MedTrinity-25M is a large-scale multimodal medical dataset featuring over 25 million images from 10 imaging modalities. It provides multigranular annota...

    work page

  71. [71]

    MMInstruct-GPT4V • Publisher: Shanghai AI Laboratory et al. • Size: 378186 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/yuecao0119/ MMInstruct-GPT4V • Description: MMInstruct-GPT4V is a multilingual multi-modal instruction tuning dataset for visual question answering and image captioning, licensed under Apache-2.0. It comprises ...

    work page

  72. [72]

    Comprised of three core components—148.4K molecule- oriented instructions (e.g

    Mol-Instructions • Publisher: Zhejiang University • Size: over 2 million instances • License: CC-BY-4.0 • Link: https://huggingface.co/datasets/zjunlp/Mol-Instructions • Description: Mol-Instructions is an open-access, large-scale biomolecular instruction dataset with 100M-1B examples designed to facilitate instruction-tuning of large language models on c...

    work page

  73. [73]

    It encompasses over one million samples in English and Chinese across five splits—helpfulness, honesty and harmlessness—totaling 2.16 GB of text

    MOSS_002_sft_data • Publisher: Fudan University • Size: 1161137 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/fnlp/moss-002-sft-data • Description: MOSS_002_sft_data is an open-source bilingual conversational dataset de- signed for fine-tuning MOSS-002. It encompasses over one million samples in English and Chinese across five ...

    work page

  74. [74]

    needles”) hidden within multi-turn conversations. Inspired by Gemini’s MRCR, it embeds 2, 4, or 8 duplicate prompts (e.g., “Write a poem about tapirs

    MRCR • Publisher: OpenAI • Size: 2400 instances • License: MIT • Link: https://huggingface.co/datasets/openai/mrcr • Description: OpenAI MRCR (Multi-round co-reference resolution) is a long-context bench- mark evaluating LLMs’ ability to find multiple identical requests (“needles”) hidden within multi-turn conversations. Inspired by Gemini’s MRCR, it embe...

    work page

  75. [75]

    NATURAL INSTRUCTIONS • Publisher: Allen Institute for AI et al. • Size: 61 datasets • License: Apache-2.0 • Link: https://huggingface.co/datasets/Muennighoff/ natural-instructions • Description: NATURAL INSTRUCTIONS is a monolingual English dataset derived from Super-Natural-Instructions, offering 1,600+ NLP tasks for training, validation, and testing. Si...

    work page

  76. [76]

    Nemotron-CrossThink • Publisher: NVIDIA • Size: 588645 instances • License: CC-BY-4.0 • Link: https://huggingface.co/datasets/nvidia/ Nemotron-CrossThink • Description: Nemotron-CrossThink is a multi-domain reinforcement learning dataset de- signed to enhance both general-purpose and mathematical reasoning in large language models. It comprises two subset...

    work page

  77. [77]

    New Y orker Caption Ranking • Publisher: University of Wisconsin-Madison et al. • Size: 2183522 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/yguooo/newyorker_ caption_ranking • Description: The New Yorker Caption Ranking dataset comprises over 250 million massive crowdsourced humor ratings on more than 2.2 million captions col...

    work page

  78. [78]

    No Robots • Publisher: Hugging Face H4 • Size: 10000 instances • License: CC-BY-NC-4.0 • Link: https://huggingface.co/datasets/HuggingFaceH4/no_robots • Description: No Robots is a high-quality, human-curated instruction dataset comprising 10,000 examples for supervised fine-tuning of language models. It includes 9,500 training and 500 test instances acro...

    work page

  79. [79]

    NuminaMath-1.5 • Publisher: Numina • Size: 896215 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/AI-MO/NuminaMath-1.5 • Description: NuminaMath-1.5 is an open-source, large-scale post-training dataset compris- ing about 900 000 competition-level mathematics problems paired with chain-of-thought solutions. It covers diverse sources...

    work page

  80. [80]

    It includes over 461,000 quality ratings and more than 10,000 fully annotated trees

    OASST1 • Publisher: OpenAssistant • Size: 161443 instances • License: Apache-2.0 • Link: https://huggingface.co/datasets/OpenAssistant/oasst1 • Description: OpenAssistant Conversations (OASST1) is a human-generated, human- annotated corpus with 161,443 messages in 66,497 conversation trees across 35 languages. It includes over 461,000 quality ratings and ...

    work page

Showing first 80 references.