The purpose of this dataset is to make it easy to convert a language model pretrained on large amounts of text into an instruction following model using a small amount of additional compute via finetuning or softprompt tuning.
These instructions are derived from open source locations or created synthetically by myself. Do your own research if you believe you may need a special permission for your application. Some of the instructions were rephrased from academic NLP datasets (see subsections for more information). I attest that the synthetically generated instructions were generated with FLAN-UL2-20B or FLAN-T5-11B models using few shot prompting so there should not be any use restrictions. Those instructions are available as a separate file for download. To the best of my knowledge none of the data in any of my datasets were derived from OpenAI APIs or services. Responses to the instructions were generated using long chat dialogue prompts that were either entirely or substantially written by myself. The generations were made from various models freely available on huggingface using my own GPUs or pre-emptible compute provided to LAION via the Stability compute cluster. Thank you Stability and LAION for your support of open source projects!
These datasets contain synthetic data and in some cases data that includes humans trying to get the language model to say toxic/offensive/trolling things. If you are concerned about the presence of this type of material in the dataset please make sure you carefully inspect each of the entries and filter appropriately. Our goal is for the model to be as helpful and non-toxic as possible and we are actively evaluating ways to reduce or eliminate undesirable content from the instruction tuning datasets.
We have seen some promising capabilities from instruction tuning with the following mix of datasets that are derived from datasets available online.
The files for this data are in json format as a list of tuples where each tuple is (prompt,agent_response) instruction_tuning_dataset_alpha_part1.json instruction_tuning_dataset_alpha_part2.json open_natural_instructions_dataset_alpha.json (this is only the synthetic instruction response examples created as described above)
A balanced set of diverse natural and factual questions and answers made using few shot prompted UL2 20B and an instruction tuned GPT-NeoX-20B model (Chip) or other models available on the huggingface hub and then rejection sampled using multiple automatic evaluations to remove low quality outputs and to filter out factually inaccurate answers. Also includes a small number of filtered natural instructions from Anthropic Helpful instructions (see: https://github.com/anthropics/hh-rlhf).
A set of filtered and reformatted instruction / response pairs where the agent response contains a list. Rejection filtered instruction response pairs generated by Chip20B or other models available on the huggingface hub that contained lists. A small portion were sourced from the Anthropic github (see: https://github.com/anthropics/hh-rlhf). Some lists were sourced from wikihow text lists created by b-mc2 (https://huggingface.co/datasets/b-mc2/wikihow_lists). All lists are formatted in a similar style.
A set of instruction / response pairs sourced from the Anthropic redteam paper github (see: https://github.com/anthropics/hh-rlhf). This dataset includes a lot of data regarding real humans trying to make the Anthropic language models say harmful/toxic/trolling things. For this dataset only examples that were rated lowly on the harmful scale (0,1,2 out of 4, where 4 is the most toxic) were included. Again, only the first lines of dialogue (instruction, first_agent_response) were retained.
Examples of instructions and responses where an appropriate response is to ask for more information from the prompter. These examples were generated from a combination of few shot prompted UL2 20B (to generate natural questions) and a large dialogue prompted language model to generate the responses containing follow-up questions.
Questions and answers generated from wikipedia articles that discuss potentially sensitive topics (flagged as potentially toxic by an early toxicity detection model).
GSM8K is a dataset of 8.5K high quality linguistically diverse grade school math word problems created by human problem writers. The dataset is segmented into 7.5K training problems and 1K test problems. These problems take between 2 and 8 steps to solve, and solutions primarily involve performing a sequence of elementary calculations using basic arithmetic operations (+ − ×÷) to reach the final answer. A bright middle school student should be able to solve every problem. It can be used for multi-step mathematical reasoning. (https://github.com/openai/grade-school-math)
Examples from the Com2Sense and Strategy QA datasets that were reformatted into natural instructions using large language models with few shot prompting and additional quality filtering steps.
Examples of instructions and responses for the generation of character or scene descriptions. Scenes were sourced from video game wikis and reformatted into instruction / response format using large language models or generated by few shot prompting with large language models.
A set of instruction / response pairs where the User requests the agent to generate a python function. These examples were generated using a large language model and few shot prompting with python code verified to execute. There are also ~3000 examples of manually curated one line python code examples from the Conala publication (see: https://conala-corpus.github.io/)