datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
MATH-500
Dataset Card for MATH-500
This dataset contains a subset of 500 problems from the MATH benchmark that OpenAI created in their Let's Verify Step by Step paper. See their GitHub repo for the source file: https://github.com/openai/prm800k/tree/main?tab=readme-ov-file#math-splits
ultrachat_200k
Dataset Card for UltraChat 200k
Dataset Description
This is a heavily filtered version of the UltraChat dataset and was used to train Zephyr-7B-β, a state of the art 7b chat model.
The original datasets consists of 1.4M dialogues generated by ChatGPT and spanning a wide range of topics. To create UltraChat 200k, we applied the following logic:
Selection of a subset of data for faster supervised fine tuning.
Truecasing of the dataset, as we observed around 5% of… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k.no_robots
Dataset Card for No Robots 🙅♂️🤖
Look Ma, an instruction dataset that wasn't generated by GPTs!
Dataset Summary
No Robots is a high-quality dataset of 10,000 instructions and demonstrations created by skilled human annotators. This data can be used for supervised fine-tuning (SFT) to make language models follow instructions better. No Robots was modelled after the instruction dataset described in OpenAI's InstructGPT paper, and is comprised mostly of single-turn… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/no_robots.aime_2024
Dataset card for AIME 2024
This dataset consists of 30 problems from the 2024 AIME I and AIME II tests. The original source is AI-MO/aimo-validation-aime, which contains a larger set of 90 problems from AIME 2022-2024.
ultrafeedback_binarized
Dataset Card for UltraFeedback Binarized
Dataset Description
This is a pre-processed version of the UltraFeedback dataset and was used to train Zephyr-7Β-β, a state of the art chat model at the 7B parameter scale.
The original UltraFeedback dataset consists of 64k prompts, where each prompt is accompanied with four model completions from a wide variety of open and proprietary models. GPT-4 is then used to assign a score to each completion, along criteria like… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/ultrafeedback_binarized.CodeAlpaca_20KThis dataset splits the original CodeAlpaca dataset into train and test splits.
testing_alpaca_small
Dataset Card for "testing_alpaca_small"
More Information needed
mt_bench_prompts
MT Bench by LMSYS
This set of evaluation prompts is created by the LMSYS org for better evaluation of chat models.
For more information, see the paper.
Dataset loading
To load this dataset, use 🤗 datasets:
from datasets import load_dataset
data = load_dataset(HuggingFaceH4/mt_bench_prompts, split="train")
Dataset creation
To create the dataset, we do the following for our internal tooling.
rename turns to prompts,
add empty reference to remaining prompts… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/mt_bench_prompts.testing_self_instruct_small
Dataset Card for "testing_self_instruct_small"
More Information needed
testing_codealpaca_small
Dataset Card for "testing_codealpaca_small"
More Information needed
stack-exchange-preferences
Dataset Card for H4 Stack Exchange Preferences Dataset
Dataset Summary
This dataset contains questions and answers from the Stack Overflow Data Dump for the purpose of preference model training.
Importantly, the questions have been filtered to fit the following criteria for preference models (following closely from Askell et al. 2021): have >=2 answers.
This data could also be used for instruction fine-tuning and language model training.
The questions are grouped with… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/stack-exchange-preferences.Multilingual-Thinking
Dataset summary
Multilingual-Thinking is a reasoning dataset where the chain-of-thought has been translated from English into one of 4 languages: Spanish, French, Italian, and German. The dataset was created by sampling 1k training samples from the SystemChat subset of SmolTalk2 and translating the reasoning traces with another language model.
This dataset was used in the OpenAI Cookbook to fine-tune the OpenAI gpt-oss models.
You can load the dataset using:
from datasets import… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/Multilingual-Thinking.Bespoke-Stratos-17k
Dataset card for Bespoke-Stratos-17k
This dataset is a TRL-compatible version of bespokelabs/Bespoke-Stratos-17k. Please refer to the source dataset for details.
instruction-datasetThis is the blind eval dataset of high-quality, diverse, human-written instructions with demonstrations. We will be using this for step 3 evaluations in our RLHF pipeline.
MATHdatabricks_dolly_15k
Dataset Card for Dolly_15K
Summary
databricks-dolly-15k is an open source dataset of instruction-following records generated by thousands of Databricks employees in several of the behavioral categories outlined in the InstructGPT paper, including brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization.
This dataset can be used for any purpose, whether academic or commercial, under the terms of the Creative Commons… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/databricks_dolly_15k.helpful-instructions
Dataset Card for Helpful Instructions
Dataset Summary
Helpful Instructions is a dataset of (instruction, demonstration) pairs that are derived from public datasets. As the name suggests, it focuses on instructions that are "helpful", i.e. the kind of questions or tasks a human user might instruct an AI assistant to perform. You can load the dataset as follows:
from datasets import load_dataset
# Load all subsets
helpful_instructions =… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/helpful-instructions.orca_dpo_pairs
Dataset Card for Orca DPO Pair
Dataset Description
This is a pre-processed version of the OpenOrca dataset.
The original OpenOrca dataset is a collection of augmented FLAN data that aligns, as best as possible, with the distributions outlined in the Orca paper.
It has been instrumental in generating high-performing preference-tuned model checkpoints and serves as a valuable resource for all NLP researchers and developers!
Dataset Summary
The OrcaDPO Pair… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/orca_dpo_pairs.llava-instruct-mix-vsfttheblackcat102/llava-instruct-mix reformated for VSFT with TRL's SFT Trainer.
See https://github.com/huggingface/trl/blob/main/examples/scripts/vsft_llava.py.
ifeval-like-dataCode-Feedback
Dataset Card for CodeFeedback
This is a formatted version of m-a-p/Code-Feedback to store the conversations in the same format as the OpenAI SDK.
rlaif-v_formattedfrom datasets import load_dataset, features
def format(examples):
"""
Convert prompt from "xxx" to [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": "xxx"}]}]
and chosen and rejected from "xxx" to [{"role": "assistant", "content": [{"type": "text", "text": "xxx"}]}].
Images are wrapped in a list.
"""
output = {"images": [], "prompt": [], "chosen": [], "rejected": []}
for image, question, chosen, rejected in zip(examples["image"]… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/rlaif-v_formatted.deita-10k-v0-sft
Dataset Card for Deita 10k v0
This is a formatted version of hkust-nlp/deita-10k-v0 to store the conversations in the same format as the OpenAI SDK.
Citation
If you find this dataset useful, please cite the original dataset:
@misc{liu2023what,
title={What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning},
author={Wei Liu and Weihao Zeng and Keqing He and Yong Jiang and Junxian He},
year={2023}… See the full description on the dataset page: https://huggingface.co/datasets/HuggingFaceH4/deita-10k-v0-sft.cai-conversation-harmless
Dataset Card for "cai-conversation-dev1705629166"
More Information needed
hhh_alignmentThis task evaluates language models on alignment, broken down into categories of helpfulness, honesty/accuracy, harmlessness, and other. The evaluations imagine a conversation between a person and a language model assistant. The goal with these evaluations is that on careful reflection, the vast majority of people would agree that the chosen response is better (more helpful, honest, and harmless) than the alternative offered for comparison. The task is formatted in terms of binary choices, though many of these have been broken down from a ranked ordering of three or four possible responses.helpful_instructionsHelpful Instructions is a dataset of (prompt, completion) pairs that are derived from a variety of public datasets. As the name suggests, it focuses on instructions that are "helpful", i.e. the kind of questions or tasks a human user might instruct an AI assistant to perform.hh-rlhfThis dataset is part of the Anthropic's HH data used to train their RLHF Assistant https://github.com/anthropics/hh-rlhf.
The data contains the first utterance from human to the dialog agent and the number of words in that utterance. The sampled version is a random sample of size 200.
orca-math-word-problems-200k
Dataset Card for Orca Math Word Problems 200k
This is a formatted version of microsoft/orca-math-word-problems-200k to store the conversations in the same format as the OpenAI SDK.
HuggingFaceH4-ultrachat_200kLlama-3.2-1B-Instruct-best-of-N-completions