To search datasets programmatically: GET https://api.databazaar.io/datasets?query=your-search
Full API docs: https://api.databazaar.io/llms.txt
Agent discovery: https://databazaar.io/.well-known/agent.json
Browse Data
145–168 of 247MS MARCO Question Answering & Passage Ranking
Bing search questions with human-generated answers and passage relevance judgments for retrieval and question answering. Supplier notes flag possible noncommercial research restrictions; review the official source terms for your use case.
- Records
- 1,112,939
- Format
- PARQUET
- Downloads
- 2
UltraChat Multi-Round Dialogue Dataset
Large-scale multi-round conversation dataset generated via dual ChatGPT Turbo API role-play, designed for instruction tuning and chat model fine-tuning.
- Records
- 773,913
- Format
- PARQUET
- Downloads
- 2
MATH-500 Benchmark
500-problem subset of the MATH benchmark used in mathematical reasoning evaluation for language models and reasoning systems. Covers diverse mathematics domains and difficulty levels.
- Records
- 500
- Format
- PARQUET
- Downloads
- 3
BigCodeBench Code Generation Benchmark
Code generation benchmark with 1,140 tasks in docstring-based completion and natural-language instruction variants, covering multiple programming languages and complexity levels.
- Records
- 5,700
- Format
- PARQUET
- Downloads
- 3
WildChat Real Human-ChatGPT Conversations
Real-world conversations between human users and ChatGPT (GPT-3.5/4), including demographics, timestamps, languages, and toxicity labels. Widely used for instruction tuning, evaluation, and alignment research.
- Records
- 837,989
- Format
- PARQUET
- Downloads
- 2
HelpSteer3 Multilingual Preference Dataset
NVIDIA's human preference and feedback dataset for reward model training and LLM alignment. Covers 15 languages with samples, achieving 85.5% on RM-Bench and 78.6% on JudgeBench.
- Records
- 132,937
- Format
- PARQUET
- Downloads
- 4
Alpaca Cleaned Instruction Fine-Tuning Dataset
Cleaned version of Stanford's Alpaca instruction-following dataset with hallucinations, merged instructions, and empty outputs removed. Suitable for large language model fine-tuning.
- Records
- 51,760
- Format
- PARQUET
- Downloads
- 1
OpenR1-Math-220k Reasoning Traces
Math problems with verified DeepSeek R1 reasoning traces sourced from NuminaMath 1.5, designed for training and evaluating mathematical reasoning models.
- Records
- 450,258
- Format
- PARQUET
- Downloads
- 3
ScienceQA Multimodal Science Questions
Multimodal multiple-choice science questions with images, lectures, and chain-of-thought explanations across natural science, social science, and language science. Widely used for vision-language model evaluation and chain-of-thought fine-tuning.
- Records
- 21,208
- Format
- PARQUET
- Downloads
- 3
Nemotron Content Safety Dataset V2
Annotated human-LLM interactions for content safety classification across 12+ harm categories, designed for training and evaluating LLM safety guardrails.
- Records
- 33,416
- Format
- PARQUET
- Downloads
- 1
Hermes Agent Reasoning Traces
Multi-turn agent tool-calling trajectories with step-by-step reasoning traces and real tool execution results, generated from Kimi-K2.5 and GLM-5.1 models via the Hermes Agent harness.
- Records
- 14,701
- Format
- PARQUET
- Downloads
- 5
UltraChat 200k Zephyr SFT Dataset
Multi-turn ChatGPT-generated conversations for supervised fine-tuning of chat models. Filtered subset used to train Zephyr-7B-β.
- Records
- 515,311
- Format
- PARQUET
- Downloads
- 3
NuminaMath-TIR Math Problems
Math problems with tool-integrated reasoning traces generated via GPT-4, derived from NuminaMath-CoT. Designed for training math-reasoning agents with code execution capability.
- Records
- 72,540
- Format
- PARQUET
- Downloads
- 2
Anthropic HH-RLHF Preference Triples
Anthropic's Helpful & Harmless RLHF dataset reformatted as prompt/chosen/rejected triples for preference modeling and DPO/RLHF training.
- Records
- 124,503
- Format
- PARQUET
- Downloads
- 3
Wikimedia Wikipedia Multilingual Corpus
Cleaned full-text Wikipedia articles across 300+ language editions from official Wikimedia dumps, formatted as one row per article. Foundational corpus for LLM pretraining, RAG, and multilingual NLP applications.
- Records
- 61,614,907
- Format
- PARQUET
- Downloads
- 3
SWE-bench GitHub Issue Resolution Benchmark
Real-world issue-pull request pairs from 12 popular Python repositories, with resolution verified by post-PR unit tests. Designed to evaluate language model and agent performance on practical software engineering tasks.
- Records
- 21,527
- Format
- PARQUET
- Downloads
- 3
C4 Colossal Clean Crawled Corpus
Cleaned Common Crawl web text corpus spanning English and 108 additional languages. Foundational pretraining dataset used in T5 and numerous open language models.
- Records
- 114,005,516
- Format
- PARQUET
- Downloads
- 3
FinePDFs Multilingual Corpus
PDF-sourced text corpus covering 1,733 languages across hundreds of millions of documents. Designed for pretraining and multilingual NLP research.
- Records
- 476,178,356
- Format
- PARQUET
- Downloads
- 2
FineWeb-Edu Educational Web Content Corpus
Educational web pages filtered from FineWeb using a Llama3-70B-trained classifier. Suitable for LLM pretraining and retrieval-augmented generation tasks.
- Records
- 3,496,736,741
- Format
- PARQUET
- Downloads
- 3
Common Corpus Multilingual Text
Open and permissibly-licensed text dataset across books, newspapers, scientific articles, legal documents, code, and other sources in 13+ languages, designed for LLM pretraining.
- Records
- 69,907
- Format
- PARQUET
- Downloads
- 3
MMLU Benchmark
Multiple-choice benchmark spanning 57 subjects including humanities, STEM, social sciences, and professional domains. Evaluates language model knowledge and reasoning across approximately 116,000 questions.
- Records
- 231,400
- Format
- PARQUET
- Downloads
- 3
FineWeb2 Multilingual Web Corpus
Filtered web text for language model pretraining across 1000+ languages. Second iteration of HuggingFace's FineWeb dataset, validated through ablation experiments.
- Records
- 4,484,929,995
- Format
- PARQUET
- Downloads
- 3
Wikipedia Embeddings BGE-M3 Multilingual
June 2024
Paragraph embeddings from Wikipedia across 11 languages, generated with the multilingual BGE-M3 model. Suitable for multilingual semantic search, retrieval-augmented generation, and retrieval evaluation tasks.
- Records
- 11,800,000
- Format
- PARQUET
- Downloads
- 3
SWE-bench Pro
Enterprise-level benchmark for evaluating AI agents on long-horizon software engineering tasks. Contains challenging real-world coding problems following SWE-Bench Verified structure.
- Records
- 731
- Format
- PARQUET
- Downloads
- 14