Datasets:

argilla
/

ultrafeedback-binarized-curation

Modalities:

Tabular

Text

Formats:

Size:

Libraries:

Dataset card Data Studio Files Files and versions

xet

Community

Dataset Viewer

Auto-converted to Parquet Duplicate

Split (1)

train · 64k rows

source stringclasses 9 values	instruction stringlengths 7 14.5k	best_rated_is_different_from_best_overall bool 2 classes	best_overall_model stringclasses 17 values	score_best_overall float64 2 10 ⌀	best_rated_model stringclasses 17 values	score_best_rated float64 0.75 5 ⌀	best_overall_score_response dict	random_response_for_best_overall dict	best_rated_response dict	random_response_for_best_rated dict	score_random_response_for_best_overall float64 1 10 ⌀	score_random_response_for_rated float64 0.75 5 ⌀	completions listlengths 0 4	random_response_for_rated float64	best_overall_score_response_critique_sentiment float64 0 1 ⌀
evol_instruct	Can you write a C++ program that prompts the user to enter the name of a country and checks if it borders the Mediterranean Sea? Here's some starter code to help you out: #include <iostream> #include <string> using namespace std; int main() { string country; // prompt user for input cout << "Enter the name ...	false	starchat	7.5	starchat	5	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and provides comprehensive information without being lengthy or repetitive.", "Rationale For Rating": "The code is accurate, in-depth, and provides a detailed solution to the task, making it outstan...	{ "annotations": { "helpfulness": { "Rating": "2", "Rationale": "The response is clear and not lengthy, but it lacks useful and comprehensive information.", "Rationale For Rating": "The code is partially incorrect as it checks if the country name ends with \"Mediterranean\" instead of checking i...	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and provides comprehensive information without being lengthy or repetitive.", "Rationale For Rating": "The code is accurate, in-depth, and provides a detailed solution to the task, making it outstan...	{ "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and provides comprehensive information without being lengthy or repetitive.", "Rationale For Rating": "The code is accurate and provides valuable insights, but it uses a vector and the find method i...	4	4	[ { "annotations": { "helpfulness": { "Rating": "2", "Rationale": "The response is clear and not lengthy, but it lacks useful and comprehensive information.", "Rationale For Rating": "The code is partially incorrect as it checks if the country name ends with \"Mediterranean\" instead...	null	0.01449
evol_instruct	Suppose you are a content creator and want to generate compelling titles and descriptions for your YouTube videos automatically. You have decided to use GPT to solve this problem. Can you provide an example of how you would use GPT to generate titles and descriptions for your YouTube videos automatically? Please includ...	false	gpt-4	7.5	gpt-4	4.75	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information and avoiding repetition.", "Rationale For Rating": "The text is accurate, in-depth, and offers profound insights and comprehensive information, includ...	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information and avoiding repetition.", "Rationale For Rating": "The text is accurate, in-depth, and offers profound insights and comprehensive information on usin...	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information and avoiding repetition.", "Rationale For Rating": "The text is accurate, in-depth, and offers profound insights and comprehensive information, includ...	{ "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information and avoiding repetition.", "Rationale For Rating": "The text is accurate and extensive, providing valuable insights and detailed information, but it l...	7.5	4.5	[ { "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information and avoiding repetition.", "Rationale For Rating": "The text is accurate, in-depth, and offers profound insights and comprehensive infor...	null	0.066424
evol_instruct	Identify the interrelated economic, political, and social factors that contributed to the stock market crash of 1929, including but not limited to the impact of World War I on the global economy, the role of government policies such as the Smoot-Hawley Tariff Act, the effects of speculative investment practices and mar...	true	mpt-30b-chat	7	vicuna-33b	4.5	{ "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information without being lengthy or repetitive.", "Rationale For Rating": "The text is highly informative, accurate, and extensive, providing valuable insights a...	{ "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information without being lengthy or repetitive.", "Rationale For Rating": "The text is highly informative, accurate, and extensive, providing valuable insights a...	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information without being lengthy or repetitive.", "Rationale For Rating": "The text is outstandingly helpful, accurate, in-depth, and provides profound insights ...	{ "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information without being lengthy or repetitive.", "Rationale For Rating": "The text is highly informative, accurate, and extensive, providing valuable insights a...	6	4.25	[ { "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear, relevant, and comprehensive, providing useful information without being lengthy or repetitive.", "Rationale For Rating": "The text is highly informative, accurate, and extensive, providing valua...	null	0.997121
evol_instruct	How can I convert the decimal number 31 to binary format using JavaScript code? Can you provide the code for this conversion?	false	mpt-30b-chat	9	mpt-30b-chat	5	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear and relevant, provides comprehensive information, and avoids repetition.", "Rationale For Rating": "The code provided is correct, and the explanation is detailed and informative.", "Type": [ "1",...	{ "annotations": { "helpfulness": { "Rating": "3", "Rationale": "The response is clear and relevant, provides comprehensive information, and avoids repetition.", "Rationale For Rating": "The code provided is correct, but it is more complex than necessary. A simpler solution using `toString(2)` w...	{ "annotations": { "helpfulness": { "Rating": "5", "Rationale": "The response is clear and relevant, provides comprehensive information, and avoids repetition.", "Rationale For Rating": "The code provided is correct, and the explanation is detailed and informative.", "Type": [ "1",...	{ "annotations": { "helpfulness": { "Rating": "2", "Rationale": "The response is clear and relevant, provides comprehensive information, and avoids repetition.", "Rationale For Rating": "The code provided is incorrect, as there is no `toBinaryString()` method in JavaScript. The correct method is...	7.5	2	[ { "annotations": { "helpfulness": { "Rating": "4", "Rationale": "The response is clear and relevant, provides useful information, and is concise.", "Rationale For Rating": "The code provided is correct and concise, but it lacks a detailed explanation of the code.", "Type": ...	null	0.999198
evol_instruct	"Can you modify the C++ code provided below to generate the first 20 Fibonacci numbers, using the fo(...TRUNCATED)	true	llama-2-13b-chat	6	ultralm-13b	4	{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is related to the task, provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"3","Rationale":"The response is related to the task, provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"3","Rationale":"The response is related to the task, provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is related to the task, provid(...TRUNCATED)	5	1.75	[{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is related to the task, provi(...TRUNCATED)	null	0.001894
evol_instruct	"A factory produces two types of toys: robots and cars. The factory can produce 50 robots and 100 ca(...TRUNCATED)	true	llama-2-70b-chat	6	bard	4.25	{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	5	2.75	[{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provi(...TRUNCATED)	null	0.059931
evol_instruct	"Explain the concept of \"quantum entanglement\" in a way that a non-scientist can understand, while(...TRUNCATED)	false	llama-2-13b-chat	7.5	llama-2-13b-chat	5	{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The text is clear and relevant, provides co(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The text is clear and relevant, provides us(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The text is clear and relevant, provides co(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The text is clear and relevant, provides us(...TRUNCATED)	7	4.5	[{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The text is clear and relevant, provides c(...TRUNCATED)	null	0.995785
evol_instruct	"Can you solve this language puzzle? Think about the impact of climate change in a certain region.\n(...TRUNCATED)	true	bard	4	gpt-4	3.25	{"annotations":{"helpfulness":{"Rating":"1","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"1","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"2","Rationale":"The response is clear and relevant to the t(...TRUNCATED)	2	1.75	[{"annotations":{"helpfulness":{"Rating":"1","Rationale":"The response is clear, relevant, and provi(...TRUNCATED)	null	0.005263
evol_instruct	"How can we enhance the robotic arm's capabilities for real-time object recognition and path plannin(...TRUNCATED)	false	gpt-3.5-turbo	8.5	gpt-3.5-turbo	5	{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear, relevant, and provid(...TRUNCATED)	7	4	[{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The response is clear, relevant, and provi(...TRUNCATED)	null	0.006974
evol_instruct	"Incorporating birth/death dates, accomplishments, and societal impact, how can you create a program(...TRUNCATED)	true	llama-2-70b-chat	7.5	gpt-3.5-turbo	4.5	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear and relevant to the t(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"5","Rationale":"The response is clear and relevant to the t(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear and relevant to the t(...TRUNCATED)	{"annotations":{"helpfulness":{"Rating":"3","Rationale":"The response is clear and relevant to the t(...TRUNCATED)	6.5	3.75	[{"annotations":{"helpfulness":{"Rating":"4","Rationale":"The response is clear and relevant to the (...TRUNCATED)	null	0.98631

End of preview. Expand in Data Studio

Ultrafeedback binarized dataset using the mean of preference ratings

Introduction

This dataset contains the result of curation work performed by Argilla (using Argilla 😃).

After visually browsing around 200 examples using the sort and filter feature of Argilla, we noticed a strong mismatch between the overall_score in the original UF dataset (and the Zephyr train_prefs dataset) and the quality of the chosen response.

By adding the critique rationale to our Argilla Dataset, we confirmed the critique rationale was highly negative, whereas the rating was very high (the highest in fact: 10). See screenshot below for one example of this issue. After some quick investigation, we identified hundreds of examples having the same issue and a potential bug on the UltraFeedback repo.

For context, this is the corresponding example within the train_prefs dataset with a score_chosen of 10.

The dataset is fully open and browsable at https://huggingface.co/spaces/argilla/ultrafeedback-curator (credentials: owner/12345678). Try browsing by discarded or using the sort feature to find problematic records yourself.

Dataset processing

We have identified a buggy behaviour of how overall_score was generated in the UltraFeedback dataset using the Critique Model, which caused very low quality (and rated) responses to get a very high score. The reason is this line which will give a 10 to responses that get a 1 from the Critique model.
To benefit from the preference data of UltraFeedback (aspect-based preference data: honesty, instruction-following, etc.) and not the Critique model (which evaluates a single response individually), we have opted for not using overall_score and compute the mean of preference ratings instead.
We select the best reponse based on this mean (named best_rated_response), and keep the one based on the overall_score for comparison purposes
We select a random response with lower mean rating (or equal in the worst case scenario, for preference tuning we'll filter those cases out), named random_response_for_best_rated. This follows the method described in the Zephyr paper of picking a random response instead of the lowest rated response. In any case, we keep all completions for people looking at additional approaches. One could binarize the data differently, for example generating several pairs per row based on their ranking (as done on the OpenAI work).
We have tried to keep all additional data for reproducibility.

Please note that *_best_overall scores are in the [1,10] range and *_best_rated are in the [1,5] range.

Based on an initial analysis, using mean rating vs overall_score picks a different chosen response in ~30K examples (out of ~63K). Additionally, using overall_score results in picking responses from less powerful models more often. See the distribution below:

Downloads last month: 294