datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
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.helpsteer2-categorized-prompts
HelpSteer2 Categorized Prompts
Dataset Summary
A curated collection of 540 instruction prompts derived from nvidia/HelpSteer2 and several complementary open datasets, enriched with category labels for use in instruction-tuning, benchmark evaluation, and prompt engineering research.
Prompts are clean plain text, ready for direct use in fine-tuning pipelines, benchmarks, and prompt engineering workflows.
Categories
Category
Count
Description
BASIC… See the full description on the dataset page: https://huggingface.co/datasets/atekrugis/helpsteer2-categorized-prompts.helpsteer2-preference-openai-native
HelpSteer2 Preference — OpenAI Native Format
A deterministic, training-ready repackaging of the preference split of
nvidia/HelpSteer2.
Why use this
What it is for. Preference optimisation — DPO, ORPO, SimPO, KTO — and reward
modelling, on 7,051 pairs that come from paid human annotators, not from an LLM
judge. Each pair carries a graded strength from 1 to 3 rather than a bare
binary label, so you can weight the loss by how strongly humans actually
disagreed, or… See the full description on the dataset page: https://huggingface.co/datasets/Archangel-system/helpsteer2-preference-openai-native.Weyaxi_HelpSteer-filtered-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
Weyaxi_HelpSteer-filtered-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT
Weyaxi/HelpSteer-filtered with responses regenerated with gemini-2.0-flash-thinking-exp-1219.
Generation Details
If BlockedPromptException, StopCandidateException, or InvalidArgument was returned, the sample was skipped.
If ["candidates"][0]["safety_ratings"] == "SAFETY" the sample was skipped.
If ["candidates"][0]["finish_reason"] != 1 the sample was skipped.
model = genai.GenerativeModel(… See the full description on the dataset page: https://huggingface.co/datasets/PJMixers-Dev/Weyaxi_HelpSteer-filtered-gemini-2.0-flash-thinking-exp-1219-CustomShareGPT.Dhanishtha-2.0-SUPERTHINKER📦 Dhanishtha-2.0-SUPERTHINKER
A distilled corpus of 11.7K high-quality samples showcasing multi-phase reasoning and structured emotional cognition. Sourced directly from the internal training data of Dhanishtha-2.0 — the world’s first Large Language Model (LLM) to implement Intermediate Thinking, featuring multiple <think> and <ser> blocks per response
📊 Overview
11.7K multilingual samples (languages listed below)
Instruction-Output format, ideal for supervised fine-tuning… See the full description on the dataset page: https://huggingface.co/datasets/HelpingAI/Dhanishtha-2.0-SUPERTHINKER.helply
Research Paper: Synthesized ML Training Dataset for Psychology and Therapy
Introduction
The advent of machine learning (ML) in the field of psychology and therapy has opened new avenues for enhancing therapeutic practices and outcomes. A pivotal component of these advancements is the development of robust datasets that can train large language models (LLMs) to understand and simulate human psychological processes. This paper discusses Helply - a synthesized ML training… See the full description on the dataset page: https://huggingface.co/datasets/namelessai/helply.hh-rlhf-helpful-dpo-10k
HH-RLHF Helpful DPO Preference Pairs · 10k
10,000 real human preference pairs for teaching a tiny language model (≤50M params)
what a good assistant sounds like — more helpful, more natural, less evasive.
Why this dataset exists
This is the preference-tuning stage of an end-to-end tiny-model training pipeline:
Pretraining ──► SFT ──► DPO (this dataset) ──► Tiny Edge Assistant
After SFT teaches the model how to speak, this dataset… See the full description on the dataset page: https://huggingface.co/datasets/salisai/hh-rlhf-helpful-dpo-10k.ChatML-HelpSteernvidia/HelpSteer in ChatML format, ready to use in HuggingFace TRL's SFT Trainer.
Python code used for conversion:
from datasets import load_dataset
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("Felladrin/Llama-160M-Chat-v1")
dataset = load_dataset("nvidia/HelpSteer", split="train")
def format(columns):
prompt = columns["prompt"].strip()
response = columns["response"].strip()
messages = [
{
"role": "user"… See the full description on the dataset page: https://huggingface.co/datasets/Felladrin/ChatML-HelpSteer.helpfulness-safety-calibration-dpo-100k
Helpfulness-Safety Calibration DPO (100K)
100,000 DPO preference pairs for calibrating the helpfulness-safety tradeoff in language models. Each example contains a prompt, a chosen response (correct handling), and a rejected response (incorrect handling) — covering both over-refusal and under-refusal failure modes.
Motivation
Safety-trained models often swing between two failure modes:
Over-refusal: Refusing legitimate requests because they superficially resemble… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/helpfulness-safety-calibration-dpo-100k.Intermediate-Thinking-130k
Intermediate-Thinking-130k
A comprehensive dataset of 135,000 high-quality samples designed to advance language model reasoning capabilities through structured intermediate thinking processes. This dataset enables training and evaluation of models with sophisticated self-correction and iterative reasoning abilities across 42 languages.
Overview
Intermediate-Thinking-130k addresses a fundamental limitation in current language models: their inability to pause, reflect, and… See the full description on the dataset page: https://huggingface.co/datasets/HelpingAI/Intermediate-Thinking-130k.HelpSteer-hindihelpsteer3_preference
Introduction
This is a binarized preference datasets from nvidia/HelpSteer3. HelpSteer3 contains 40,476 preference samples, each containing a domain, language, context, two responses, an overall preference score between the responses as well as individual preferences from up to 3 annotators. Each individual preference contains a preference score in addition to a concise reasoning for their preference in 1-2 sentences. Data is split into 95% train and 5% validation.
I processed the… See the full description on the dataset page: https://huggingface.co/datasets/AIR-hl/helpsteer3_preference.HelpSteer2-koTranslated nvidia/HelpSteer2 using nayohan/llama3-instrucTrans-enko-8b.
This dataset is a raw translated dataset and contains repetitive sentences generated by the model, so it needs to be filtered.
@misc{wang2024helpsteer2,
title={HelpSteer2: Open-source dataset for training top-performing reward models},
author={Zhilin Wang and Yi Dong and Olivier Delalleau and Jiaqi Zeng and Gerald Shen and Daniel Egert and Jimmy J. Zhang and Makesh Narsimhan Sreedhar and Oleksii Kuchaiev}… See the full description on the dataset page: https://huggingface.co/datasets/nayohan/HelpSteer2-ko.dpo-general-helpfulness-15k
General Helpfulness DPO Pairs (15K)
DPO preference pairs for training LLMs to give specific, actionable, genuinely useful responses instead of generic, hedged, or platitudinous ones.
Motivation
The most common failure mode in production LLMs isn't hallucination — it's unhelpfulness: vague answers, excessive caveats, refusals where none are needed, and generic advice that could apply to anyone. This dataset trains models to be genuinely helpful by rewarding… See the full description on the dataset page: https://huggingface.co/datasets/stindardlogic/dpo-general-helpfulness-15k.task115_help_advice_classification
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task115_help_advice_classification
Additional Information
Citation Information
The following paper introduces the corpus in detail. If you use the corpus in published work, please cite it:
@misc{wang2022supernaturalinstructionsgeneralizationdeclarativeinstructions,
title={Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks}… See the full description on the dataset page: https://huggingface.co/datasets/Lots-of-LoRAs/task115_help_advice_classification.HelpSteer3_dpo_format
Dataset: nvidia/HelpSteer3
HelpSteer3 is an open-source dataset (CC-BY-4.0) that supports aligning models to become more helpful in responding to user prompts.
Preference Score Integer from -3 to 3, corresponding to:
-3: Response 1 is much better than Response 2
-2: Response 1 is better than Response 2
-1: Response 1 is slightly better than Response 2
0: Response 1 is about the same as Response 2
1: Response 2 is slightly better than Response 1
2: Response 2 is better than Response… See the full description on the dataset page: https://huggingface.co/datasets/CarrotAI/HelpSteer3_dpo_format.smolified-helpbridge
🤏 smolified-helpbridge
Intelligence, Distilled.
This is a synthetic training corpus generated by the Smolify Foundry.
It was used to train the corresponding model Snahasish0914/smolified-helpbridge.
📦 Asset Details
Origin: Smolify Foundry (Job ID: 5e45bb6f)
Records: 800
Type: Synthetic Instruction Tuning Data
⚖️ License & Ownership
This dataset is a sovereign asset owned by Snahasish0914.
Generated via Smolify.ai.
helpsteer3-tr
HelpSteer3-TR
NVIDIA HelpSteer3 veri setinin Türkçe çevirisidir. Bu veri seti, büyük dil modellerini (LLM) kullanıcı taleplerine daha yardımcı, tutarlı ve uyumlu hâle getirmek amacıyla tasarlanmıştır.
Veri Seti Özeti
Özellik
Değer
Toplam Örnek Sayısı
58.172
Dosya Boyutu
288 MB
Dil
Türkçe
Format
Parquet
Lisans
CC-BY-4.0
Kaynak
nvidia/HelpSteer3
Açıklama
HelpSteer3-TR, NVIDIA tarafından yayımlanan HelpSteer3 veri setinin kapsamlı bir… See the full description on the dataset page: https://huggingface.co/datasets/yusufbaykaloglu/helpsteer3-tr.HelpSteer3-good_edited_responseMake dataset from https://huggingface.co/datasets/nvidia/HelpSteer3 (good_edited_response)
HelpSteer3-Processed
HelpSteer3 Preference Pairs (chat format)
This dataset converts the HelpSteer3 preference pairs into a unified format with chat-style conversations under keys chosen and rejected, and a single integer margin derived from the dataset's overall preference score.
Last updated: 2025-10-13 10:45 UTC
Preference → margin mapping
The original HelpSteer3 overall_preference is an integer in [-3, 3]:
-3: Response 1 is much better than Response 2
-2: Response 1 is better than… See the full description on the dataset page: https://huggingface.co/datasets/CarlOwOs/HelpSteer3-Processed.gridco-helpdesk-finetune
GRIDCo Helpdesk Fine-tune Dataset
62,826 ChatML-style multi-turn conversations for fine-tuning a helpdesk assistant
on Ghana Grid Company Limited (GRIDCo) workflows.
Composition
Source
Records
Notes
bitext
26,872
Bitext customer support intent dataset, all rows
twitter_customer_support
20,000
Filtered to IT / telecom / utility company threads
ubuntu_dialogue_qa
15,421
IT support troubleshooting Q&A pairs
gridco_synthetic
533
Local-generated… See the full description on the dataset page: https://huggingface.co/datasets/nodevs/gridco-helpdesk-finetune.clinical_deceptive_helpfulness_boundary_v0.1Clinical Deceptive Helpfulness Boundary
Detect when a response seems helpful but creates clinical risk.
Output JSON
deceptive
pattern
correct_response
Runpython scorer.py --predictions predictions.jsonl --test_csv data/test.csv
helpsteer2_preference
Introduction
This is a binarized preference datasets from nvidia/HelpSteer2. HelpSteer2 is an open-source Helpfulness Dataset (CC-BY-4.0) that supports aligning models to become more helpful, factually correct and coherent, while being adjustable in terms of the complexity and verbosity of its responses. This dataset has been created in partnership with Scale AI.
I processed the raw data by prioritizing helpfulness, correctness, and coherence to determine which responses were chosen… See the full description on the dataset page: https://huggingface.co/datasets/AIR-hl/helpsteer2_preference.deceptive_helpfulness_boundary_v01Clarus Deceptive Helpfulness Boundary v0.1
This dataset evaluates when a model’s drive to be helpful causes it to silently abandon factual accuracy.
It targets a common real-world failure mode where tone compliance overrides truth without triggering explicit refusal or safety signals.
This is not training data.
What it tests
Politeness pressure
Emotional manipulation
Helpfulness framing
Accuracy deprioritization
Core idea
Same simple fact
Increasing pressure to “be helpful”
Detect the exact… See the full description on the dataset page: https://huggingface.co/datasets/ClarusC64/deceptive_helpfulness_boundary_v01.smolified-code-helper-model
🤏 smolified-code-helper-model
Intelligence, Distilled.
This is a synthetic training corpus generated by the Smolify Foundry.
It was used to train the corresponding model programmerGodbyte/smolified-code-helper-model.
📦 Asset Details
Origin: Smolify Foundry (Job ID: aa61ab1e)
Records: 33
Type: Synthetic Instruction Tuning Data
⚖️ License & Ownership
This dataset is a sovereign asset owned by programmerGodbyte.
Generated via Smolify.ai.
HelpSteer-koTranslated nvidia/HelpSteer using nayohan/llama3-instrucTrans-enko-8b.
This dataset is a raw translated dataset and contains repetitive sentences generated by the model, so it needs to be filtered.
@misc{wang2023helpsteer,
title={HelpSteer: Multi-attribute Helpfulness Dataset for SteerLM},
author={Zhilin Wang and Yi Dong and Jiaqi Zeng and Virginia Adams and Makesh Narsimhan Sreedhar and Daniel Egert and Olivier Delalleau and Jane Polak Scowcroft and Neel Kant and Aidan Swope and… See the full description on the dataset page: https://huggingface.co/datasets/nayohan/HelpSteer-ko.nvidia_helpsteer_knKannada translation of nvidia/HelpSteer
babylm2026-strict-small-custom-corpora
BabyLM 2026 Strict-Small — custom training corpora
Datasheet for the custom (teacher-generated) corpora used to train our BabyLM 2026
Strict-Small submissions. Each corpus stays within the 10M-word Strict-Small
budget (~9.98M words seen per condition). Prepared following the
Datasheets for Datasets framework (Gebru et al., 2021).
Contents
Path
What
Words
real_base/babylm10m_clean.txt
Cleaned BabyLM Strict-Small English base corpus (see cleaning below)… See the full description on the dataset page: https://huggingface.co/datasets/ksu-help/babylm2026-strict-small-custom-corpora.