[
  {
    "instruction": "Given a raw dataset metadata record, identify whether it is suitable for supervised fine-tuning (SFT) of LLMs and summarize its best use case.",
    "input": "{\"source\":\"huggingface\",\"id\":\"fka/prompts.chat\",\"name\":\"prompts.chat\",\"description\":\"a.k.a. Awesome ChatGPT Prompts. This is a Dataset Repository mirror of prompts.chat — a social platform for AI prompts. prompts.chat is an open-source platform where users can share, discover, and collect AI prompts from the community.\",\"tags\":[\"task_categories:question-answering\",\"task_categories:text-generation\",\"license:cc0-1.0\",\"size_categories:1K<n<10K\",\"format:csv\",\"modality:text\",\"library:datasets\",\"library:pandas\",\"library:polars\",\"library:mlcroissant\",\"region:us\",\"ChatGPT\",\"prompts\",\"AI\",\"GPT\",\"Claude\",\"Gemini\",\"Llama\",\"Mistral\",\"LLM\",\"prompt-engineering\",\"conversational-ai\",\"text-generation\",\"chatbot\",\"awesome-list\"],\"downloads\":50433,\"likes\":9680,\"created_at\":\"\"}",
    "output": "This dataset is only partially suitable for SFT. It appears to be a collection of community-shared prompts rather than prompt-response training pairs. Its best use case is prompt engineering research, instruction template discovery, agent prompting experiments, and synthetic data generation. For SFT, it would likely require augmentation with high-quality responses or conversion into prompt-completion examples.",
    "source": "huggingface",
    "original_id": "fka/prompts.chat"
  },
  {
    "instruction": "Given a raw dataset metadata record, identify whether it is suitable for supervised fine-tuning (SFT) of LLMs and summarize its best use case.",
    "input": "{\"source\":\"huggingface\",\"id\":\"HuggingFaceFW/fineweb\",\"name\":\"fineweb\",\"description\":\"FineWeb consists of more than 18.5T tokens of cleaned and deduplicated English web data from CommonCrawl. The data processing pipeline is optimized for LLM performance.\",\"tags\":[\"task_categories:text-generation\",\"language:en\",\"license:odc-by\",\"size_categories:10B<n<100B\",\"modality:tabular\",\"modality:text\",\"arxiv:2306.01116\",\"arxiv:2109.07445\",\"arxiv:2406.17557\",\"doi:10.57967/hf/2493\",\"region:us\"],\"downloads\":652595,\"likes\":2770,\"created_at\":\"\"}",
    "output": "This dataset is not ideal for supervised fine-tuning because it is a large-scale cleaned web corpus rather than an instruction-response dataset. Its best use case is base model pretraining or continued pretraining on high-quality English web text. It may support unsupervised language modeling, domain adaptation, and data mixture construction, but it is not directly structured for instruction SFT.",
    "source": "huggingface",
    "original_id": "HuggingFaceFW/fineweb"
  },
  {
    "instruction": "Given a raw dataset metadata record, identify whether it is suitable for supervised fine-tuning (SFT) of LLMs and summarize its best use case.",
    "input": "{\"source\":\"huggingface\",\"id\":\"Anthropic/hh-rlhf\",\"name\":\"hh-rlhf\",\"description\":\"This repository provides access to human preference data about helpfulness and harmlessness from Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback. These data are meant to train preference or reward models for subsequent RLHF training and are not meant for supervised training of dialogue agents.\",\"tags\":[\"license:mit\",\"size_categories:100K<n<1M\",\"format:json\",\"modality:text\",\"library:datasets\",\"library:dask\",\"library:mlcroissant\",\"library:polars\",\"arxiv:2204.05862\",\"region:us\",\"human-feedback\"],\"downloads\":35080,\"likes\":1722,\"created_at\":\"\"}",
    "output": "This dataset is generally not suitable for direct SFT because it is designed for preference modeling and RLHF, not supervised dialogue training. Its best use case is training reward models, preference models, or benchmarking alignment methods around helpfulness and harmlessness. If used for SFT, extreme care is needed because the dataset documentation explicitly warns against training dialogue agents directly on it.",
    "source": "huggingface",
    "original_id": "Anthropic/hh-rlhf"
  },
  {
    "instruction": "Given a raw dataset metadata record, identify whether it is suitable for supervised fine-tuning (SFT) of LLMs and summarize its best use case.",
    "input": "{\"source\":\"huggingface\",\"id\":\"Open-Orca/OpenOrca\",\"name\":\"OpenOrca\",\"description\":\"The OpenOrca dataset is a rich collection of augmented FLAN data aligned as closely as possible with the distributions outlined in the Orca paper. It has been instrumental in generating high-performing model checkpoints.\",\"tags\":[\"task_categories:text-classification\",\"task_categories:token-classification\",\"task_categories:table-question-answering\",\"task_categories:question-answering\",\"task_categories:zero-shot-classification\",\"task_categories:summarization\",\"task_categories:feature-extraction\",\"task_categories:text-generation\",\"language:en\",\"license:mit\",\"size_categories:1M<n<10M\",\"format:parquet\",\"modality:text\",\"library:datasets\",\"library:dask\",\"library:polars\",\"library:mlcroissant\",\"arxiv:2306.02707\",\"arxiv:2301.13688\",\"arxiv:2302.13971\",\"region:us\"],\"downloads\":34091,\"likes\":1524,\"created_at\":\"\"}",
    "output": "This dataset is well suited for supervised fine-tuning. It is an instruction-oriented corpus derived from augmented FLAN-style data and was created specifically to support high-performing instruction-following models. Its best use case is general-purpose instruction tuning for assistant behavior across QA, summarization, reasoning, classification, and related NLP tasks.",
    "source": "huggingface",
    "original_id": "Open-Orca/OpenOrca"
  },
  {
    "instruction": "Given a raw dataset metadata record, identify whether it is suitable for supervised fine-tuning (SFT) of LLMs and summarize its best use case.",
    "input": "{\"source\":\"huggingface\",\"id\":\"OpenAssistant/oasst1\",\"name\":\"oasst1\",\"description\":\"OpenAssistant Conversations Dataset (OASST1) is a human-generated, human-annotated assistant-style conversation corpus with 161,443 messages in 35 languages, annotated with quality ratings and structured as conversation trees.\",\"tags\":[\"language:en\",\"language:es\",\"language:ru\",\"language:de\",\"language:pl\",\"language:th\",\"language:vi\",\"language:sv\",\"language:bn\",\"language:da\",\"language:he\",\"language:it\",\"language:fa\",\"language:sk\",\"language:id\",\"language:nb\",\"language:el\",\"language:nl\",\"language:hu\",\"language:eu\",\"language:zh\",\"language:eo\",\"language:ja\",\"language:ca\",\"language:cs\",\"language:bg\",\"language:fi\",\"language:pt\",\"language:tr\",\"language:ro\",\"language:ar\",\"language:uk\",\"language:gl\",\"language:fr\",\"language:ko\",\"license:apache-2.0\",\"size_categories:10K<n<100K\",\"format:parquet\",\"modality:tabular\",\"modality:text\",\"library:datasets\",\"library:pandas\",\"library:mlcroissant\",\"library:polars\",\"arxiv:2304.07327\",\"region:us\",\"human-feedback\"],\"downloads\":11994,\"likes\":1502,\"created_at\":\"\"}",
    "output": "This dataset is well suited for supervised fine-tuning, especially for conversational assistants. It contains human-generated and human-annotated multi-turn assistant conversations with quality signals, making it valuable for chat SFT, multilingual assistant training, and response ranking or filtering. Its best use case is building multilingual instruction-following and dialogue models with strong conversational structure.",
    "source": "huggingface",
    "original_id": "OpenAssistant/oasst1"
  },
  {
    "instruction": "Given a dataset metadata record, summarize the dataset for supervised fine-tuning with a focus on purpose, modality, language, likely use cases, and notable properties.",
    "input": "{\"source\":\"huggingface\",\"id\":\"openai/gsm8k\",\"name\":\"gsm8k\",\"description\":\"Dataset Card for GSM8K. GSM8K (Grade School Math 8K) is a dataset of 8.5K high quality linguistically diverse grade school math word problems. The dataset was created to support the task of question answering on basic mathematical problems that require multi-step reasoning. These problems take between 2 and 8 steps to solve. Solutions primarily involve performing a sequence of elementary calculations using basic arithmetic operations.\",\"tags\":[\"benchmark:official\",\"benchmark:eval-yaml\",\"task_categories:text-generation\",\"annotations_creators:crowdsourced\",\"language_creators:crowdsourced\",\"multilinguality:monolingual\",\"source_datasets:original\",\"language:en\",\"license:mit\",\"size_categories:10K<n<100K\",\"format:parquet\",\"modality:text\",\"library:datasets\",\"library:pandas\",\"library:polars\",\"library:mlcroissant\",\"arxiv:2110.14168\",\"region:us\",\"math-word-problems\"],\"downloads\":823505,\"likes\":1274,\"created_at\":\"\"}",
    "output": "GSM8K is an English text dataset of grade school math word problems designed for multi-step reasoning and question answering. It contains about 8.5K high-quality problems that typically require 2 to 8 reasoning steps and basic arithmetic operations. It is commonly used as an official benchmark for evaluating mathematical reasoning in language models. The dataset is monolingual, originally created, MIT-licensed, and well suited for training or evaluating chain-of-thought style reasoning on math word problems.",
    "source": "huggingface",
    "original_id": "openai/gsm8k"
  },
  {
    "instruction": "Given a dataset metadata record, summarize the dataset for supervised fine-tuning with a focus on purpose, modality, language, likely use cases, and notable properties.",
    "input": "{\"source\":\"huggingface\",\"id\":\"wikimedia/wikipedia\",\"name\":\"wikipedia\",\"description\":\"Dataset Card for Wikimedia Wikipedia. Wikipedia dataset containing cleaned articles of all languages. The dataset is built from the Wikipedia dumps with one subset per language, each containing a single train split. Each example contains the content of one full Wikipedia article with cleaning to strip markdown and unwanted sections.\",\"tags\":[\"task_categories:text-generation\",\"task_categories:fill-mask\",\"task_ids:language-modeling\",\"task_ids:masked-language-modeling\",\"license:cc-by-sa-3.0\",\"license:gfdl\",\"size_categories:10M<n<100M\",\"format:parquet\",\"modality:text\",\"library:datasets\",\"library:dask\",\"library:polars\",\"library:mlcroissant\",\"region:us\"],\"downloads\":135291,\"likes\":1200,\"created_at\":\"\"}",
    "output": "Wikimedia Wikipedia is a massive multilingual text dataset built from cleaned Wikipedia article dumps. Each example contains a full article, with markdown and unwanted sections removed, and the dataset is organized into one subset per language. It is useful for language modeling, masked language modeling, pretraining, retrieval, and general knowledge tasks. The dataset is text-only, very large in scale, and licensed under CC-BY-SA-3.0 and GFDL, making it a common resource for broad-coverage multilingual model training.",
    "source": "huggingface",
    "original_id": "wikimedia/wikipedia"
  },
  {
    "instruction": "Given a dataset metadata record, summarize the dataset for supervised fine-tuning with a focus on purpose, modality, language, likely use cases, and notable properties.",
    "input": "{\"source\":\"huggingface\",\"id\":\"gsdf/EasyNegative\",\"name\":\"EasyNegative\",\"description\":\"Negative Embedding. This is a Negative Embedding trained with Counterfeit. Please use it in the stable-diffusion-webui embeddings folder. It can be used with other models, but the effectiveness is not certain.\",\"tags\":[\"license:other\",\"size_categories:n<1K\",\"format:imagefolder\",\"modality:image\",\"library:datasets\",\"library:mlcroissant\",\"region:us\"],\"downloads\":36062,\"likes\":1174,\"created_at\":\"\"}",
    "output": "EasyNegative is an image-related resource centered on a negative embedding intended for use with Stable Diffusion style workflows rather than a conventional supervised dataset. It appears to be a very small image-modality artifact associated with prompt conditioning or generation quality control, especially in stable-diffusion-webui setups. The licensing is nonstandard ('other'), and the description suggests practical compatibility with specific models such as Counterfeit, while effectiveness with other models may vary.",
    "source": "huggingface",
    "original_id": "gsdf/EasyNegative"
  },
  {
    "instruction": "Given a dataset metadata record, summarize the dataset for supervised fine-tuning with a focus on purpose, modality, language, likely use cases, and notable properties.",
    "input": "{\"source\":\"huggingface\",\"id\":\"togethercomputer/RedPajama-Data-1T\",\"name\":\"RedPajama-Data-1T\",\"description\":\"RedPajama is a clean-room, fully open-source implementation of the LLaMa dataset.\",\"tags\":[\"task_categories:text-generation\",\"language:en\",\"size_categories:1M<n<10M\",\"modality:text\",\"library:datasets\",\"library:mlcroissant\",\"region:us\"],\"downloads\":2127,\"likes\":1152,\"created_at\":\"\"}",
    "output": "RedPajama-Data-1T is a large English text dataset intended for text generation and large language model pretraining. It is described as a clean-room, fully open-source implementation of the LLaMA training dataset, making it relevant for open foundation model research and reproduction efforts. Its primary use cases include pretraining, continued pretraining, and large-scale corpus construction for generative language models.",
    "source": "huggingface",
    "original_id": "togethercomputer/RedPajama-Data-1T"
  },
  {
    "instruction": "Given a dataset metadata record, summarize the dataset for supervised fine-tuning with a focus on purpose, modality, language, likely use cases, and notable properties.",
    "input": "{\"source\":\"huggingface\",\"id\":\"FreedomIntelligence/medical-o1-reasoning-SFT\",\"name\":\"medical-o1-reasoning-SFT\",\"description\":\"News: [2025/04/22] We split the data and kept only the medical SFT dataset (medical_o1_sft.json). The file medical_o1_sft_mix.json contains a mix of medical and general instruction data. [2025/02/22] We released the distilled dataset from Deepseek-R1 based on medical verifiable problems. [2024/12/25] We open-sourced the medical reasoning dataset for SFT, built on medical verifiable problems and an LLM.\",\"tags\":[\"task_categories:question-answering\",\"task_categories:text-generation\",\"language:en\",\"language:zh\",\"license:apache-2.0\",\"size_categories:10K<n<100K\",\"format:json\",\"modality:text\",\"library:datasets\",\"library:pandas\",\"library:mlcroissant\",\"library:polars\",\"arxiv:2412.18925\",\"region:us\",\"medical\",\"biology\"],\"downloads\":7421,\"likes\":1087,\"created_at\":\"\"}",
    "output": "medical-o1-reasoning-SFT is a bilingual English-Chinese text dataset for supervised fine-tuning on medical reasoning and question answering. It is built around medical verifiable problems and includes reasoning-oriented SFT data, with references to distilled reasoning traces derived from DeepSeek-R1. The dataset is particularly suitable for training medical assistants, diagnostic reasoning models, and domain-specific instruction-following systems. It uses JSON format, is Apache-2.0 licensed, and spans medical and biology-related use cases.",
    "source": "huggingface",
    "original_id": "FreedomIntelligence/medical-o1-reasoning-SFT"
  },
  {
    "instruction": "Summarize the dataset for language model pretraining use, including its scale, domain, language, license, notable metadata, and likely use cases.",
    "input": "Name: fineweb-edu\nDescription: FineWeb-Edu is a 1.3 trillion token educational web dataset filtered from FineWeb using an educational quality classifier. It is the 1.3T token version and is described in arXiv:2406.17557.\nTags: task_categories:text-generation, language:en, license:odc-by, size_categories:1B<n<10B, format:parquet, modality:tabular, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2406.17557, arxiv:2404.14219, arxiv:2401.10020, arxiv:2109.07445, doi:10.57967/hf/2497, region:us\nDownloads: 361619\nLikes: 1039",
    "output": "fineweb-edu is a large-scale English educational web dataset intended for text generation and language model pretraining. It contains 1.3 trillion tokens curated from the FineWeb corpus using an educational quality classifier, making it especially suitable for training models on high-quality instructional and knowledge-rich text. The dataset is licensed under ODC-By and distributed in Parquet format with text and tabular modalities, with support across common data tooling such as Hugging Face Datasets, Dask, and Polars. Its scale, educational focus, and strong community adoption suggest it is well suited for pretraining, continued pretraining, and data mixture construction for general-purpose or education-oriented LLMs.",
    "source": "huggingface",
    "original_id": "HuggingFaceFW/fineweb-edu"
  },
  {
    "instruction": "Summarize the dataset for language model pretraining use, including its scale, domain, language, license, notable metadata, and likely use cases.",
    "input": "Name: dolma\nDescription: Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research\nTags: task_categories:text-generation, language:en, license:odc-by, size_categories:n>1T, arxiv:2402.00159, arxiv:2301.13688, region:us, language-modeling, casual-lm, llm\nDownloads: 3113\nLikes: 1021",
    "output": "dolma is an open English corpus designed explicitly for language model pretraining research, with a reported scale of three trillion tokens. It is tagged for text generation and language modeling and licensed under ODC-By, which makes it relevant for open research and large-scale LLM training pipelines. Given its massive size and explicit positioning as a pretraining corpus, it is likely best suited for foundation model pretraining, ablation studies on data mixtures, and research on scaling laws, corpus composition, and training efficiency.",
    "source": "huggingface",
    "original_id": "allenai/dolma"
  },
  {
    "instruction": "Summarize the dataset for code model training use, including its contents, scale, language coverage, licensing notes, format, and likely use cases.",
    "input": "Name: the-stack\nDescription: The Stack is a code dataset whose initial release included 30 programming languages and 18 permissive licenses, with a near-deduplicated size of 3TB. Later versions excluded weak copyleft licenses and expanded the permissive license list.\nTags: task_categories:text-generation, language_creators:crowdsourced, language_creators:expert-generated, multilinguality:multilingual, language:code, license:other, size_categories:100M<n<1B, format:parquet, modality:tabular, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2211.15533, arxiv:2107.03374, arxiv:2207.14157, region:us\nDownloads: 12610\nLikes: 985",
    "output": "the-stack is a multilingual code dataset built for text generation and code model training. It aggregates source code across many programming languages and was released with a strong focus on license filtering and near-deduplication. The dataset is distributed in Parquet format and supports text and tabular access through common data libraries. Because it covers many languages and emphasizes permissive licensing in later releases, it is well suited for pretraining code LLMs, code completion systems, repository-level modeling research, and studies of code data governance and deduplication.",
    "source": "huggingface",
    "original_id": "bigcode/the-stack"
  },
  {
    "instruction": "Summarize the dataset for small-language or curriculum-style model training, including its content style, generation method, scale, license, and likely use cases.",
    "input": "Name: TinyStories\nDescription: Dataset containing synthetically generated short stories created by GPT-3.5 and GPT-4 using a small vocabulary. Described in arXiv:2305.07759.\nTags: task_categories:text-generation, language:en, license:cdla-sharing-1.0, size_categories:1M<n<10M, format:parquet, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2305.07759, region:us\nDownloads: 94124\nLikes: 963",
    "output": "TinyStories is an English synthetic text dataset made up of short stories generated by GPT-3.5 and GPT-4 with deliberately simple vocabulary. Its relatively small size and constrained linguistic complexity make it especially useful for training compact language models, studying emergence and learning dynamics, and building curriculum-style datasets for early-stage language acquisition in models. The dataset is licensed under CDLA-Sharing-1.0, distributed in Parquet format, and is a strong fit for efficient experimentation, toy-scale pretraining, and interpretability-oriented language modeling research.",
    "source": "huggingface",
    "original_id": "roneneldan/TinyStories"
  },
  {
    "instruction": "Summarize the dataset for instruction tuning, including its task types, origin, scale, license, format, and likely use cases.",
    "input": "Name: databricks-dolly-15k\nDescription: An open source dataset of instruction-following records generated by thousands of Databricks employees across categories such as brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization.\nTags: task_categories:question-answering, task_categories:summarization, language:en, license:cc-by-sa-3.0, size_categories:10K<n<100K, format:json, modality:text, library:datasets, library:pandas, library:mlcroissant, library:polars, arxiv:2203.02155, region:us\nDownloads: 33439\nLikes: 953",
    "output": "databricks-dolly-15k is an English instruction-tuning dataset containing around 15,000 human-generated instruction-response examples created by Databricks employees. It spans a range of behavioral categories including brainstorming, classification, question answering, generation, information extraction, and summarization. The dataset is licensed under CC-BY-SA-3.0 and distributed in JSON format, making it easy to use in supervised fine-tuning pipelines. It is well suited for instruction following, chat-style supervised fine-tuning, benchmarking response behavior across task categories, and creating lightweight alignment datasets for open models.",
    "source": "huggingface",
    "original_id": "databricks/databricks-dolly-15k"
  },
  {
    "instruction": "Summarize the dataset and highlight its main purpose, modality, language, license, approximate size, and notable metadata.",
    "input": "Name: alpaca\nSource: huggingface\nID: tatsu-lab/alpaca\nDescription: Alpaca is a dataset of 52,000 instructions and demonstrations generated by OpenAI's text-davinci-003 engine. This instruction data can be used to conduct instruction-tuning for language models and make the language model follow instruction better. The authors built on the data generation pipeline from Self-Instruct framework.\nTags: task_categories:text-generation, language:en, license:cc-by-nc-4.0, size_categories:10K<n<100K, format:parquet, modality:text, library:datasets, library:pandas, library:polars, library:mlcroissant, region:us, instruction-finetuning\nDownloads: 89040\nLikes: 948",
    "output": "Alpaca is an English text dataset on Hugging Face designed for instruction fine-tuning and text generation. It contains about 52,000 instructions and demonstrations generated with text-davinci-003, following the Self-Instruct pipeline. The dataset is licensed under CC-BY-NC-4.0, stored in Parquet format, and falls in the 10K to 100K size range. It is intended to help language models better follow instructions and is commonly used with datasets, pandas, polars, and mlcroissant tooling. It has strong community adoption, with 89,040 downloads and 948 likes.",
    "source": "huggingface",
    "original_id": "tatsu-lab/alpaca"
  },
  {
    "instruction": "Summarize the dataset and highlight its main purpose, modality, language, license, approximate size, and notable metadata.",
    "input": "Name: bad_prompt\nSource: huggingface\nID: Nerfgun3/bad_prompt\nDescription: Negative Embedding / Textual Inversion. The idea behind this embedding was to train the negative prompt as an embedding, unifying the basis of the negative prompt into one word or embedding. The embedding is described as helpful for image generation, especially hands. Usage involves downloading the file and placing it in the stable-diffusion-webui embeddings folder.\nTags: language:en, license:creativeml-openrail-m, size_categories:n<1K, format:imagefolder, modality:image, library:datasets, library:mlcroissant, region:us, stable-diffusion, text-to-image, image-to-image\nDownloads: 2107\nLikes: 938",
    "output": "bad_prompt is a small English image-related dataset/resource on Hugging Face associated with Stable Diffusion workflows. It focuses on negative embedding or textual inversion, aiming to compress negative prompt behavior into a reusable embedding token. The dataset uses the imagefolder format, has fewer than 1,000 items, and is licensed under CreativeML OpenRAIL-M. It is relevant to both text-to-image and image-to-image generation, and is intended for use with Stable Diffusion WebUI embeddings. Despite relatively modest downloads at 2,107, it has 938 likes, indicating strong niche interest.",
    "source": "huggingface",
    "original_id": "Nerfgun3/bad_prompt"
  },
  {
    "instruction": "Summarize the dataset and list its key characteristics for model training use.",
    "input": "Name: falcon-refinedweb\nSource: huggingface\nDataset ID: tiiuae/falcon-refinedweb\nDescription: Falcon RefinedWeb is a massive English web dataset built by TII and released under an ODC-By 1.0 license. It is built through stringent filtering and large-scale deduplication of CommonCrawl. Models trained on RefinedWeb reportedly achieve performance in-line or better than models trained on curated datasets while relying only on web data. It is also multimodal-friendly and contains links and alt text.\nTags: task_categories:text-generation, language:en, license:odc-by, size_categories:100M<n<1B, format:parquet, modality:text, library:datasets, library:dask, library:mlcroissant, library:polars, arxiv:2306.01116, arxiv:2203.15556, arxiv:2107.06499, arxiv:2104.08758, arxiv:2109.07445, arxiv:1911.00359, arxiv:2112.11446, doi:10.57967/hf/0737, region:us\nDownloads: 40476\nLikes: 905",
    "output": "Falcon RefinedWeb is a large English web text dataset from TII designed for text generation training. It is derived from CommonCrawl using stringent filtering and large-scale deduplication, with the goal of providing high-quality web-only training data. The dataset is described as multimodal-friendly because it preserves elements such as links and alt text. Key characteristics include: language English, modality text, parquet format, size category between 100M and 1B samples, license ODC-By 1.0, and compatibility with datasets, dask, polars, and mlcroissant. It is a popular resource on Hugging Face with 40,476 downloads and 905 likes, and is associated with multiple research references including doi:10.57967/hf/0737.",
    "source": "huggingface",
    "original_id": "tiiuae/falcon-refinedweb"
  },
  {
    "instruction": "Summarize the dataset and list its key characteristics for conversational AI research.",
    "input": "Name: lmsys-chat-1m\nSource: huggingface\nDataset ID: lmsys/lmsys-chat-1m\nDescription: LMSYS-Chat-1M is a large-scale real-world LLM conversation dataset containing one million conversations with 25 state-of-the-art LLMs. It was collected from 210K unique IP addresses on the Vicuna demo and Chatbot Arena website from April to August 2023. Each sample includes a conversation ID, model name, conversation text in OpenAI API JSON format, a detected language tag, and an OpenAI moderation API tag. User consent is obtained through the Terms of use.\nTags: size_categories:1M<n<10M, format:parquet, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2309.11998, region:us\nDownloads: 8116\nLikes: 875",
    "output": "LMSYS-Chat-1M is a real-world conversational dataset for LLM research containing one million conversations involving 25 state-of-the-art language models. The data was collected from 210,000 unique IP addresses through the Vicuna demo and Chatbot Arena between April and August 2023. Each record includes a conversation ID, model name, conversation text in OpenAI API JSON format, a detected language tag, and a moderation tag. Key characteristics include: text modality, parquet format, size category between 1M and 10M samples, and support for datasets, dask, polars, and mlcroissant. It is useful for conversational AI analysis, dialogue modeling, and benchmark research, and has 8,116 downloads and 875 likes on Hugging Face.",
    "source": "huggingface",
    "original_id": "lmsys/lmsys-chat-1m"
  },
  {
    "instruction": "Summarize the dataset based on its metadata and description.",
    "input": "Name: ShareGPT_Vicuna_unfiltered\nSource: huggingface\nID: anon8231489123/ShareGPT_Vicuna_unfiltered\nDescription: Further cleaning done. Please look through the dataset and ensure that I didn't miss anything. Update: Confirmed working method for training the model: https://huggingface.co/AlekseyKorshuk/vicuna-7b/discussions/4#64346c08ef6d5abefe42c12c Two choices: Removes instances of \"I'm sorry, but\": https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/blob/main/ShareGPT_V3_unfiltered_cleaned_split_no_imsorry.json Has instances of \"I'm sorry, but\":… See the full description on the dataset page: https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered.\nTags: language:en, license:apache-2.0, region:us\nDownloads: 160837\nLikes: 861\nCreated at: ",
    "output": "ShareGPT_Vicuna_unfiltered is a Hugging Face dataset containing cleaned ShareGPT/Vicuna-style conversation data in English. The description indicates additional cleaning was performed and references a confirmed training method for Vicuna-style models. It also mentions two dataset variants: one with instances of the phrase \"I'm sorry, but\" removed and another retaining them. The dataset is tagged with Apache-2.0 licensing and a US region tag. It has strong community engagement, with 160837 downloads and 861 likes.",
    "source": "huggingface",
    "original_id": "anon8231489123/ShareGPT_Vicuna_unfiltered"
  },
  {
    "instruction": "Summarize the dataset based on its metadata and description.",
    "input": "Name: PhysicalAI-Autonomous-Vehicles\nSource: huggingface\nID: nvidia/PhysicalAI-Autonomous-Vehicles\nDescription: PHYSICAL AI AUTONOMOUS VEHICLES The PhysicalAI-Autonomous-Vehicles dataset provides one of the largest, geographically diverse collections of multi-sensor data empowering AV researchers to build the next generation of Physical AI based end-to-end driving systems. This dataset is ready for commercial/non-commercial AV use per the license agreement. Data Collection Method Automatic/Sensor Labeling Method Automatic/Sensor This dataset has a total of 1700 hours of driving… See the full description on the dataset page: https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles.\nTags: license:other, region:us\nDownloads: 386875\nLikes: 852\nCreated at: ",
    "output": "PhysicalAI-Autonomous-Vehicles is a Hugging Face dataset from NVIDIA for autonomous vehicle research. It is described as one of the largest geographically diverse multi-sensor driving datasets, intended to support development of Physical AI and end-to-end driving systems. The metadata states that both data collection and labeling were performed using automatic or sensor-based methods. The dataset includes a total of 1700 hours of driving data and is available for commercial and non-commercial AV use under its license agreement. It is tagged with a non-standard license and US region metadata, and has 386875 downloads and 852 likes.",
    "source": "huggingface",
    "original_id": "nvidia/PhysicalAI-Autonomous-Vehicles"
  },
  {
    "instruction": "Summarize the dataset for supervised fine-tuning metadata extraction.",
    "input": "Name: finepdfs\nSource: huggingface\nID: HuggingFaceFW/finepdfs\nDescription: Liberating 3T of the finest tokens from PDFs. FinePDFs is the largest publicly available corpus sourced exclusively from PDFs, containing about 3 trillion tokens across 475 million documents in 1733 languages.\nTags: task_categories:text-generation, license:odc-by, size_categories:100M<n<1B, format:parquet, modality:tabular, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2506.18421, arxiv:2109.07445, region:us\nDownloads: 19883\nLikes: 851",
    "output": "{\"dataset_name\":\"finepdfs\",\"provider\":\"HuggingFaceFW\",\"summary\":\"A massive multilingual text-generation corpus extracted exclusively from PDFs, with about 3 trillion tokens across 475 million documents in 1733 languages.\",\"task_category\":\"text-generation\",\"modalities\":[\"text\",\"tabular\"],\"format\":\"parquet\",\"license\":\"odc-by\",\"size_category\":\"100M<n<1B\",\"libraries\":[\"datasets\",\"dask\",\"polars\",\"mlcroissant\"],\"region\":\"us\",\"papers\":[\"2506.18421\",\"2109.07445\"],\"downloads\":19883,\"likes\":851,\"special_notes\":[\"PDF-only corpus\",\"multilingual\"]}",
    "source": "huggingface",
    "original_id": "HuggingFaceFW/finepdfs"
  },
  {
    "instruction": "Summarize the dataset for supervised fine-tuning metadata extraction.",
    "input": "Name: OpenThoughts-114k\nSource: huggingface\nID: open-thoughts/OpenThoughts-114k\nDescription: Open synthetic reasoning dataset with 114k high-quality examples covering math, science, code, and puzzles. Default subset contains ready-to-train data used to finetune OpenThinker-7B and OpenThinker-32B models.\nTags: license:apache-2.0, size_categories:100K<n<1M, format:parquet, modality:text, library:datasets, library:dask, library:polars, library:mlcroissant, arxiv:2506.04178, region:us, curator, synthetic\nDownloads: 142507\nLikes: 833",
    "output": "{\"dataset_name\":\"OpenThoughts-114k\",\"provider\":\"open-thoughts\",\"summary\":\"A synthetic reasoning dataset with 114k high-quality training examples spanning math, science, code, and puzzles, designed for model finetuning.\",\"task_category\":\"reasoning\",\"domains\":[\"math\",\"science\",\"code\",\"puzzles\"],\"modality\":[\"text\"],\"format\":\"parquet\",\"license\":\"apache-2.0\",\"size_category\":\"100K<n<1M\",\"libraries\":[\"datasets\",\"dask\",\"polars\",\"mlcroissant\"],\"region\":\"us\",\"papers\":[\"2506.04178\"],\"downloads\":142507,\"likes\":833,\"special_notes\":[\"synthetic\",\"ready-to-train\",\"curator viewer available\"]}",
    "source": "huggingface",
    "original_id": "open-thoughts/OpenThoughts-114k"
  },
  {
    "instruction": "Summarize the dataset for instruction tuning use.",
    "input": "Name: OpenHermes-2.5\nSource: huggingface\nID: teknium/OpenHermes-2.5\nDescription: This is the dataset that made OpenHermes 2.5 and Nous Hermes 2 series of models. It is an exact compilation and curation of many open source datasets and custom created synthetic datasets.\nTags: language:eng, size_categories:1M<n<10M, format:json, modality:text, library:datasets, library:pandas, library:mlcroissant, library:polars, region:us, synthetic, GPT-4, Distillation, Compilation\nDownloads: 23107\nLikes: 820",
    "output": "OpenHermes-2.5 is a large English text dataset hosted on Hugging Face, used to build OpenHermes 2.5 and Nous Hermes 2 model series. It consists of a curated compilation of open-source and synthetic data, with signals of GPT-4 distillation and dataset compilation. The dataset is provided in JSON format, supports common data libraries, and appears suitable for large-scale instruction tuning and synthetic data research.",
    "source": "huggingface",
    "original_id": "teknium/OpenHermes-2.5"
  },
  {
    "instruction": "Summarize the dataset for instruction tuning use.",
    "input": "Name: alpaca-cleaned\nSource: huggingface\nID: yahma/alpaca-cleaned\nDescription: This is a cleaned version of the original Alpaca Dataset released by Stanford. It fixes issues such as hallucination-prone instructions that referenced internet data.\nTags: task_categories:text-generation, language:en, license:cc-by-4.0, size_categories:10K<n<100K, format:json, modality:text, library:datasets, library:pandas, library:mlcroissant, library:polars, region:us, instruction-finetuning\nDownloads: 31316\nLikes: 813",
    "output": "alpaca-cleaned is an English instruction-tuning dataset on Hugging Face derived from Stanford's original Alpaca release. It was cleaned to remove problematic or hallucination-inducing examples, especially instructions that depended on unavailable internet data. The dataset is JSON-based, licensed under CC-BY-4.0, sized between 10K and 100K examples, and is well suited for text generation and instruction fine-tuning tasks.",
    "source": "huggingface",
    "original_id": "yahma/alpaca-cleaned"
  },
  {
    "instruction": "Summarize the dataset based on the provided metadata.",
    "input": "Name: hle\nSource: huggingface\nID: cais/hle\nDescription: Humanity's Last Exam (HLE) is a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. It consists of 2,500 questions across dozens of subjects.\nTags: benchmark:official, license:mit, size_categories:1K<n<10K, format:parquet, modality:image, modality:text, library:datasets, library:pandas, library:polars, library:mlcroissant, region:us\nDownloads: 49120\nLikes: 786",
    "output": "hle is an official Hugging Face benchmark dataset from cais/hle. It is a multimodal academic benchmark called Humanity's Last Exam, designed to test frontier-level human knowledge with broad subject coverage. The dataset contains 2,500 closed-ended questions across many subjects and supports both image and text modalities. It is distributed in parquet format under the MIT license and can be used with datasets, pandas, polars, and mlcroissant. The dataset is associated with the US region and has 49,120 downloads and 786 likes.",
    "source": "huggingface",
    "original_id": "cais/hle"
  },
  {
    "instruction": "Summarize the dataset based on the provided metadata.",
    "input": "Name: fineweb-2\nSource: huggingface\nID: HuggingFaceFW/fineweb-2\nDescription: FineWeb2 is the second iteration of the FineWeb dataset, providing high-quality pretraining data for over 1000 languages. It is fully reproducible, available under the permissive ODC-By 1.0 license, and validated through hundreds of ablation experiments.\nTags: task_categories:text-generation, license:odc-by, size_categories:1B<n<10B, modality:tabular, modality:text, arxiv:2506.20920, arxiv:2109.07445, arxiv:2406.17557, doi:10.57967/hf/3744, region:us, plus extensive language coverage across 1000+ languages\nDownloads: 119148\nLikes: 784",
    "output": "fineweb-2 is a Hugging Face dataset from HuggingFaceFW focused on large-scale multilingual text generation and pretraining. It is the second iteration of FineWeb and offers high-quality, fully reproducible text data covering more than 1,000 languages. The dataset is released under the ODC-By license, includes text and tabular modalities, and is very large in scale with a size category between 1B and 10B. It is supported by multiple research references including arXiv papers and a DOI, is associated with the US region, and has 119,148 downloads and 784 likes.",
    "source": "huggingface",
    "original_id": "HuggingFaceFW/fineweb-2"
  },
  {
    "instruction": "Given a raw dataset catalog entry, summarize it into a concise machine-learning dataset profile including its purpose, modality, likely tasks, language coverage, license, scale, and notable popularity signals.",
    "input": "{\"source\":\"huggingface\",\"id\":\"ILSVRC/imagenet-1k\",\"name\":\"imagenet-1k\",\"description\":\"ILSVRC 2012, commonly known as 'ImageNet' is an image dataset organized according to the WordNet hierarchy. Each meaningful concept in WordNet, possibly described by multiple words or word phrases, is called a synset. ImageNet aims to provide on average 1000 images to illustrate each synset.\",\"tags\":[\"task_categories:image-classification\",\"task_ids:multi-class-image-classification\",\"annotations_creators:crowdsourced\",\"language_creators:crowdsourced\",\"multilinguality:monolingual\",\"source_datasets:original\",\"language:en\",\"license:other\",\"size_categories:1M<n<10M\",\"format:parquet\",\"format:optimized-parquet\",\"modality:image\",\"library:datasets\",\"library:dask\",\"library:polars\",\"library:mlcroissant\",\"arxiv:1409.0575\",\"arxiv:1912.07726\",\"arxiv:1811.12231\",\"arxiv:2109.13228\",\"region:us\"],\"downloads\":113222,\"likes\":782,\"created_at\":\"\"}",
    "output": "{\"summary\":\"ImageNet-1k is a large-scale image dataset for multi-class image classification, based on the ILSVRC 2012 subset of ImageNet and organized by WordNet synsets.\",\"modality\":\"image\",\"primary_tasks\":[\"image classification\",\"multi-class image classification\"],\"language\":\"en\",\"multilinguality\":\"monolingual\",\"license\":\"other\",\"scale\":\"1M<n<10M\",\"data_source\":\"original\",\"annotation_type\":\"crowdsourced\",\"formats\":[\"parquet\",\"optimized-parquet\"],\"libraries\":[\"datasets\",\"dask\",\"polars\",\"mlcroissant\"],\"region\":\"us\",\"popularity\":{\"downloads\":113222,\"likes\":782},\"notes\":[\"Widely used benchmark for computer vision.\",\"Contains roughly 1000 classes represented by WordNet synsets.\"]}",
    "source": "huggingface",
    "original_id": "ILSVRC/imagenet-1k"
  },
  {
    "instruction": "Given a raw dataset catalog entry, summarize it into a concise machine-learning dataset profile including its purpose, modality, likely tasks, language coverage, license, scale, and notable popularity signals.",
    "input": "{\"source\":\"huggingface\",\"id\":\"QingyiSi/Alpaca-CoT\",\"name\":\"Alpaca-CoT\",\"description\":\"Instruction-Finetuning Dataset Collection (Alpaca-CoT). This repository continuously collects various instruction tuning datasets and standardizes them into the same format for direct use with Alpaca-style models. It also includes chain-of-thought related data.\",\"tags\":[\"language:en\",\"language:zh\",\"language:ml\",\"license:apache-2.0\",\"region:us\",\"Instruction\",\"Cot\"],\"downloads\":9734,\"likes\":755,\"created_at\":\"\"}",
    "output": "{\"summary\":\"Alpaca-CoT is a collection of standardized instruction-tuning datasets designed for Alpaca-style model fine-tuning, with emphasis on instruction following and chain-of-thought style data.\",\"modality\":\"text\",\"primary_tasks\":[\"instruction tuning\",\"supervised fine-tuning\",\"chain-of-thought reasoning\"],\"languages\":[\"en\",\"zh\",\"ml\"],\"license\":\"apache-2.0\",\"region\":\"us\",\"dataset_type\":\"collection\",\"popularity\":{\"downloads\":9734,\"likes\":755},\"notes\":[\"Standardizes multiple instruction datasets into a unified format.\",\"Useful for training or fine-tuning instruction-following language models.\"]}",
    "source": "huggingface",
    "original_id": "QingyiSi/Alpaca-CoT"
  },
  {
    "instruction": "Summarize the dataset and identify its main purpose, modalities, languages, supported task categories, license, and notable characteristics.",
    "input": "Dataset name: PersonaHub\nSource: huggingface\nID: proj-persona/PersonaHub\nDescription: Scaling Synthetic Data Creation with 1,000,000,000 Personas. This repo releases data introduced in the paper \"Scaling Synthetic Data Creation with 1,000,000,000 Personas.\" It proposes a persona-driven data synthesis methodology that leverages various perspectives within a large language model to create diverse synthetic data. PERSONA HUB is a collection of 1 billion diverse personas automatically curated from web data.\nTags: task_categories:text-generation, task_categories:text-classification, task_categories:token-classification, task_categories:fill-mask, task_categories:table-question-answering, language:en, language:zh, license:cc-by-nc-sa-4.0, size_categories:100K<n<1M, format:json, modality:text, library:datasets, library:dask, library:mlcroissant, library:polars, arxiv:2406.20094, region:us, synthetic, text, math, reasoning, instruction, tool, persona\nDownloads: 7966\nLikes: 744",
    "output": "PersonaHub is a Hugging Face dataset focused on scaling synthetic data creation through a persona-driven methodology. Its main purpose is to support diverse synthetic data generation by providing a large collection of automatically curated personas derived from web data, as described in the paper \"Scaling Synthetic Data Creation with 1,000,000,000 Personas\" (arXiv:2406.20094).\n\nKey properties:\n- Modalities: text\n- Languages: English and Chinese\n- Task categories: text generation, text classification, token classification, fill-mask, and table question answering\n- Format: JSON\n- License: CC-BY-NC-SA-4.0\n- Libraries: datasets, dask, mlcroissant, polars\n- Region: US\n\nNotable characteristics:\n- Synthetic dataset\n- Persona-focused resource for instruction, reasoning, math, and tool-related use cases\n- High community interest, with 7,966 downloads and 744 likes\n- Although the description highlights 1 billion personas, the tagged size category is 100K<n<1M, which may reflect the published dataset partition or repository metadata rather than the full conceptual scale described in the paper.",
    "source": "huggingface",
    "original_id": "proj-persona/PersonaHub"
  }
]