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
25–48 of 104Go-Code-Large Corpus
Curated collection of Go source code samples in JSONL format, designed for LLM pretraining, code generation fine-tuning, and static analysis research on cloud-native and backend systems.
- Records
- 316,427
- Format
- PARQUET
- Downloads
- 3
TaskTrove Agentic Tasks
Collection of agentic tasks aggregated from 100+ sources including SWE-Smith, R2EGym, and SWE-Re-Bench, designed for reinforcement learning and supervised fine-tuning of AI agents.
- Records
- 17,191
- Format
- PARQUET
- Downloads
- 3
Bhasha SFT Multilingual Instruction-Response Pairs
Instruction-response pairs in Hindi, Bengali, Gujarati, and English combining human-annotated and synthetic data for supervised fine-tuning of multilingual language models.
- Records
- 18,139,035
- Format
- PARQUET
- Downloads
- 1
SciCode Scientific Domain Code
Scientific source code from GitHub spanning biology, chemistry, materials science, physics, and 174 additional domains. Covers 155,855 code repositories with 1.1 billion+ lines of code across approximately 115 GB.
- Records
- 155,855
- Format
- PARQUET
- Downloads
- 1
HPDv3 Human Preference Dataset
Text-to-image preference dataset comprising 1.08M text-image pairs and 1.17M pairwise human annotations. Used to train HPSv3 reward models for evaluating generated images.
- Records
- 1,154,324
- Format
- PARQUET
- Downloads
- 3
LongCoT Long-Horizon Reasoning Benchmark
Benchmark dataset for evaluating sustained chain-of-thought reasoning across logic, computer science, chemistry, chess, and mathematics domains.
- Records
- 5,004
- Format
- PARQUET
- Downloads
- 7
UGMathBench Undergraduate Math Reasoning Benchmark
Undergraduate-level math problems spanning 16 subjects and 111 topics with 10 answer types. Designed for evaluating LLM mathematical reasoning capabilities.
- Records
- 5,061
- Format
- PARQUET
- Downloads
- 2
Common Crawl Creative Commons Fine Multilingual Corpus
Filtered Creative Commons web corpus from Common Crawl, intersected with FineWeb for quality. Covers English, German, French, Dutch, Spanish, Italian, Afrikaans, and Frisian.
- Records
- 75,055,472
- Format
- PARQUET
- Downloads
- 3
MASSIVE Multilingual NLU Dataset
Parallel natural language understanding benchmark with utterances annotated across 60 intents and 55 slot types, covering 51 languages. Localized from voice assistant interactions.
- Records
- 2,560,755
- Format
- PARQUET
- Downloads
- 2
Salad-Data LLM Safety Evaluation Benchmark
Aggregated safety questions for LLM red-teaming and jailbreak evaluation, sourced from HH-RLHF, AdvBench, ToxicChat, GPTFuzzer, and GPT-3.5 self-instructed prompts.
- Records
- 30,358
- Format
- PARQUET
- Downloads
- 3
CodeParrot Clean — Deduplicated Python Code
Cleaned and deduplicated Python source code from GitHub, filtered for line length, alphanumeric fraction, and auto-generated content. Commonly used for code language model pretraining and fine-tuning.
- Records
- 477,028
- Format
- PARQUET
- Downloads
- 3
MASC: Massive Arabic Speech Corpus
Multi-dialect Arabic speech audio at 16 kHz with transcripts, sourced from YouTube channels. Suitable for Arabic automatic speech recognition, text-to-speech, and speech language model training and evaluation.
- Records
- 913,400
- Format
- PARQUET
- Downloads
- 2
LLM-jp-4 Thinking Japanese Reasoning SFT Dataset
Supervised fine-tuning dataset pairing Japanese prompts from diverse sources with step-by-step reasoning traces and final responses for chain-of-thought model training.
- Records
- 3,199,986
- Format
- PARQUET
- Downloads
- 2
MAPS Multilingual Agentic AI Benchmark
Benchmark combining performance and security evaluation tasks across 11 languages, integrating GAIA, SWE-bench, and MATH datasets with agentic security assessments.
- Records
- 9,680
- Format
- PARQUET
- Downloads
- 1
Falcon RefinedWeb Corpus
English web corpus derived from CommonCrawl through stringent filtering and deduplication. Comprises approximately 968 million documents used to train the Falcon language model family.
- Records
- 968,000,015
- Format
- PARQUET
- Downloads
- 3
GSM8K Grade School Math Word Problems
Linguistically diverse grade-school math word problems with step-by-step solutions, used as a standard benchmark for multi-step arithmetic reasoning in language models.
- Records
- 17,584
- Format
- PARQUET
- Downloads
- 3
OpenWebMath Corpus
Mathematical documents filtered from Common Crawl, suitable for LLM math pretraining and fine-tuning. Includes diverse mathematical web text across multiple domains.
- Records
- 6,315,233
- Format
- PARQUET
- Downloads
- 2
Nemotron-Math-v2 Mathematical Reasoning Trajectories
Collection of 347K mathematical problems paired with 7M model-generated reasoning trajectories for training mathematical reasoning. Includes long-context reasoning, tool-use integration, and multi-mode supervision signals.
- Records
- 7,085,839
- Format
- PARQUET
- Downloads
- 2
OpenOrca Augmented FLAN Instruction Dataset
GPT-4 and GPT-3.5 augmented instruction-response pairs aligned with the Orca paper distribution, designed for instruction tuning and fine-tuning open language models.
- Records
- 2,942,029
- Format
- PARQUET
- Downloads
- 4
SmolLM Corpus Pretraining Data
Educational and synthetic pretraining corpus combining Cosmopedia v2 synthetic textbooks and stories with Python-Edu and deduplicated FineWeb-Edu subsets, curated for small language model training.
- Records
- 236,980,453
- Format
- PARQUET
- Downloads
- 4
Nemotron Safety Guard Dataset v3 Multilingual
Multilingual safety dataset across 12 languages for training LLM safety guard models, generated via the CultureGuard pipeline.
- Records
- 514,617
- Format
- PARQUET
- Downloads
- 3
StackMathQA Mathematics Question-Answer Corpus
Mathematical questions and answers from Stack Exchange sites including Math, MathOverflow, and Stats. Suitable for math reasoning fine-tuning, pretraining, and retrieval-augmented generation tasks.
- Records
- 6,195,432
- Format
- PARQUET
- Downloads
- 3
WildChat-4.8M — Filtered Conversation Subset
Real user–ChatGPT conversations collected by AI2, filtered using automated toxicity classifiers. Moderation and PII redaction are imperfect; this subset is not guaranteed safe for all audiences.
- Records
- 3,199,860
- Format
- PARQUET
- Downloads
- 2
TriviaQA Reading Comprehension QA
Question-answer-evidence triples for open-domain and reading comprehension QA. Widely used benchmark for retrieval-augmented QA, RAG evaluation, and extractive/abstractive QA model training.
- Records
- 847,579
- Format
- PARQUET
- Downloads
- 2