[
  {
    "path": "LICENSE",
    "content": "MIT License\n\nCopyright (c) 2023 PolisAI\n\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\n\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\n\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n"
  },
  {
    "path": "README.md",
    "content": "# Awesome Human Preference Datasets for LLM 🧑❤️🤖\nA curated list of open source **Human Preference** datasets for LLM instruction-tuning, RLHF and evaluation.\n\nFor general NLP datasets and text corpora, check out [this](https://github.com/niderhoff/nlp-datasets) awesome list.\n\n\n## Datasets\n[**OpenAI WebGPT Comparisons**](https://huggingface.co/datasets/openai/webgpt_comparisons)\n- 20k comparisons where each example comprises a question, a pair of model answers, and human-rated preference scores for each answer. \n- RLHF dataset used to train the [OpenAI WebGPT](https://arxiv.org/abs/2112.09332) reward model.\n\n[**OpenAI Summarization**](https://huggingface.co/datasets/openai/summarize_from_feedback)\n- 64k text summarization examples including human-written responses and human-rated model responses. \n- RLHF dataset used in the [OpenAI Learning to Summarize from Human Feedback](https://arxiv.org/abs/2009.01325) paper.\n- Explore sample data [here](https://openaipublic.blob.core.windows.net/summarize-from-feedback/website/index.html#/tldr_comparisons).\n\n[**Anthropic Helpfulness and Harmlessness Dataset (HH-RLHF)**](https://huggingface.co/datasets/Anthropic/hh-rlhf) \n- In total 170k human preference comparisons, including human preference data collected for [Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback](https://arxiv.org/pdf/2204.05862.pdf) and human-generated red teaming data from [Red Teaming Language Models to Reduce Harms](https://arxiv.org/abs/2209.07858), divided into 3 sub-datasets:\n    - A **base** dataset using a context-distilled 52B model, with 44k helpfulness comparisons and 42k red-teaming (harmlessness) comparisons.\n    - A **RS** dataset of 52k helpfulness comparisons and 2k red-teaming comparisons using rejection sampling models, where rejection sampling used a preference model trained on the base dataset.\n    - An iterated **online** dataset including data from RLHF models, updated weekly over five weeks, with 22k helpfulness comparisons.\n\n[**OpenAssistant Conversations Dataset (OASST1)**](https://huggingface.co/datasets/OpenAssistant/oasst1)\n- A human-generated, human-annotated assistant-style conversation corpus consisting of 161k messages in 35 languages, annotated with 461k quality ratings, resulting in 10k+ fully annotated conversation trees. \n\n[**Stanford Human Preferences Dataset (SHP)**](https://huggingface.co/datasets/stanfordnlp/SHP) \n- 385K collective human preferences over responses to questions/instructions in 18 domains for training RLHF reward models and NLG evaluation models. Datasets collected from Reddit.\n\n[**Reddit ELI5**](https://huggingface.co/datasets/eli5)\n- 270k examples of questions, answers and scores collected from 3 Q&A subreddits.\n\n[**Human ChatGPT Comparison Corpus (HC3)**](https://huggingface.co/datasets/Hello-SimpleAI/HC3)\n- 60k human answers and 27K ChatGPT answers for around 24K questions.\n- Sibling dataset available for [Chinese](https://huggingface.co/datasets/Hello-SimpleAI/HC3-Chinese).\n\n[**HuggingFace H4 StackExchange Preference Dataset**](https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences)\n- 10 million questions (with >= 2 answers) and answers (scored based on vote count) from Stackoverflow. \n\n[**ShareGPT.com**](https://sharegpt.com/)\n- 90k (as of April 2023) user-uploaded ChatGPT interactions.\n- ~~To access the data using ShareGPT's API, see documentation [here](https://github.com/domeccleston/sharegpt#rest-api)~~ The ShareGPT API is currently disabled (\"due to excess traffic\"). \n- [Precompliled datasets](https://huggingface.co/datasets?sort=downloads&search=sharegpt) on HuggingFace.\n\n[**Alpaca**](https://huggingface.co/datasets/tatsu-lab/alpaca)\n- 52k instructions and demonstrations generated by OpenAI's text-davinci-003 engine for _self-instruct_ training.\n\n[**GPT4All**](https://huggingface.co/datasets/nomic-ai/gpt4all_prompt_generations)\n- 1M prompt-response pairs colleced using GPT-3.5-Turbo API in March 2023. [GitHub repo](https://github.com/nomic-ai/gpt4all).\n\n[**Databricks Dolly Dataset**](https://huggingface.co/datasets/databricks/databricks-dolly-15k)\n- 15k instruction-following records generated [by Databricks employees](https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm) in categories including brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization.\n\n[**HH_golden**](https://huggingface.co/datasets/Unified-Language-Model-Alignment/Anthropic_HH_Golden)\n- 42k harmless data, same prompts and \"rejected\" responses as the Harmless dataset in [Anthropic HH datasets](https://huggingface.co/datasets/Anthropic/hh-rlhf), but the responses in the \"chosen\" responses are re-writtened using GPT4 to yield more harmless answers. The comparison before and after re-written can be found [here](https://huggingface.co/datasets/Unified-Language-Model-Alignment/Anthropic_HH_Golden). Empirically, compared with the original Harmless dataset, training on this dataset improves the harmless metrics for various alignment methods such as RLHF and DPO.\n"
  }
]