datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
discoverybenchData-driven Discovery Benchmark from the paper:
"DiscoveryBench: Towards Data-Driven Discovery with Large Language Models"
🔭 Overview
DiscoveryBench is designed to systematically assess current model capabilities in data-driven discovery tasks and provide a useful resource for improving them. Each DiscoveryBench task consists of a goal and dataset(s). Solving the task requires both statistical analysis and semantic reasoning. A faceted evaluation allows open-ended… See the full description on the dataset page: https://huggingface.co/datasets/allenai/discoverybench.agent-llm-traces
Multi-Benchmark LLM Agent Traces
A comprehensive dataset of OpenTelemetry traces capturing LLM inference behavior across multiple agent frameworks, benchmarks, and model providers. This dataset enables research into LLM performance analysis, agent behavior patterns, and inference optimization.
Collected by Exgentic - A platform for LLM observability and performance optimization.
Dataset Overview
This dataset contains 1,781 execution traces capturing detailed agent… See the full description on the dataset page: https://huggingface.co/datasets/DiscoPosse/agent-llm-traces.discoverybench
DiscoveryBench - Alias
A reformatted version of the original DiscoveryBench dataset for easier usage.
🤗 Original Dataset on HF
💻 GitHub Repository
📄 Paper (arXiv)
📁 Dataset Structure
The dataset consists of real and synthetic subsets:
Real Splits:
real_train
real_test
Synthetic Splits:
synth_train
synth_dev
synth_test
Each split contains a list of tasks with references to associated CSV datasets needed to answer the query. LLMs are expected to use the… See the full description on the dataset page: https://huggingface.co/datasets/nhop/discoverybench.suno-discord-chat-history
Suno Discord Community Dataset
Dataset Description
This dataset contains messages exported from the official Suno Discord server, covering multiple channels across feedback, announcements, community hubs, and the Suno Studio product space. It captures authentic user interactions, feature requests, bug reports, and community discussions around Suno's AI music generation platform.
Quick Start
Installation
pip install pandas huggingface_hub datasets… See the full description on the dataset page: https://huggingface.co/datasets/hafizhrafizal/suno-discord-chat-history.lmcache-agentic-traces
LMCache Agentic Dataset Collection
A curated dataset collection of 787 multi-turn agentic LLM sessions (24,881 total LLM iterations) designed for benchmarking stateful LLM serving systems. Every session exhibits at least 5 turns with prefix growth and builds to at least 10K tokens of context — making it ideal for evaluating tiered KV Cache solutions like LMCache.
Motivation
Modern LLM agents (coding assistants, research agents, tool-calling systems) make dozens of… See the full description on the dataset page: https://huggingface.co/datasets/DiscoPosse/lmcache-agentic-traces.red-pill-drug-discovery-formulation
🔴 RED-PILL
Research Enhanced Dataset for Pharmaceutical Innovation in Learning & Language
The first open instruction-tuning dataset for drug discovery & formulation development.
Built for fine-tuning Heretic-ablated models that won't refuse your pharmaceutical R&D questions.
⚡ Quick Start
from datasets import load_dataset
# Load the full dataset
ds = load_dataset("saidutta69/red-pill-drug-discovery-formulation"… See the full description on the dataset page: https://huggingface.co/datasets/saidutta69/red-pill-drug-discovery-formulation.discussllm
DiscussLLM
DiscussLLM is a synthetic dataset for the "when to speak" setting in multi-party
discussions. Each example contains a scenario, a discussion transcript, and one
Nexus assistant intervention.
The release contains 88,718 generated discussions with the original split:
train: 75,411 discussions
test: 13,307 discussions
Codebase: https://github.com/necla-ml/DiscussLLM
Citation
@article{patel2025discussllm,
title={DiscussLLM: Teaching Large Language Models… See the full description on the dataset page: https://huggingface.co/datasets/deepsworld/discussllm.byob-pd-book-corpus
Dataset Card for BYOB LM on Steroids - Public-Domain Book Training Corpus
A multilingual, public-domain corpus of literary, philosophical, and scientific
works, published for character-level language-model pre-training. It ships in two
forms: raw per-author .txt files (one file per author) and a pre-tokenized,
memmap-ready cache (train.bin / val.bin / meta.json) for fast training.
Dataset Sources
Project Gutenberg (https://www.gutenberg.org) - all tiers
Standard… See the full description on the dataset page: https://huggingface.co/datasets/disco-jack-basement/byob-pd-book-corpus.task563_discofuse_answer_generation
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task563_discofuse_answer_generation
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/task563_discofuse_answer_generation.task550_discofuse_sentence_generation
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task550_discofuse_sentence_generation
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/task550_discofuse_sentence_generation.task564_discofuse_classification
Dataset Card for Natural Instructions (https://github.com/allenai/natural-instructions) Task: task564_discofuse_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/task564_discofuse_classification.divergent-discourses-tibetan-newspapers
Divergent Discourses — Early Tibetan Newspapers, 1950–1965
523,215 text regions from early Tibetan-language newspapers published between 1950 and 1965,
produced by the Divergent Discourses project (SOAS University of London and Leipzig
University, with Trinity College Dublin).
This is not a flat text dump. Each row is one text region from a scanned page, retaining
its reading-order position, region type, source newspaper, and issue date — so page structure
survives the… See the full description on the dataset page: https://huggingface.co/datasets/biglam/divergent-discourses-tibetan-newspapers.discover-and-prove
MiniF2F-Hard & FIMO-Hard
Expert-reannotated Hard Mode variants of the MiniF2F and FIMO theorem-proving
benchmarks, released with our paper Discover and Prove: An Open-source Agentic
Framework for Hard Mode Automated Theorem Proving in Lean 4 (ACL 2026).
In Hard Mode, the final answer is not embedded in the formal statement:
the system must first discover the answer before constructing a formal proof —
mirroring what a human competitor actually faces. Each solution-style… See the full description on the dataset page: https://huggingface.co/datasets/liuchengwu/discover-and-prove.discursos-senado-legislatura-56
Discursos da 56ª Legislatura do Senado Federal
Visão geral
Corpus de pronunciamentos do Plenário do Senado Federal relativos à 56ª Legislatura (2019–2023), coletados da API pública do Senado e consolidados em Parquet e CSV. Cada registro corresponde a um pronunciamento, com metadados e texto integral quando disponibilizado pela fonte.
Versão atual
Esta documentação corresponde à versão v1.1.1. Os arquivos de dados são os mesmos da versão v1.1.0; a… See the full description on the dataset page: https://huggingface.co/datasets/fabriciosantana/discursos-senado-legislatura-56.RUCAIBox-Story-Generation-Alpacahttps://huggingface.co/datasets/RUCAIBox/Story-Generation
RUC AI Box HC Story Generation augmented and converted to alpaca format.
No filtering has been done.
discord-archive
Discord Archive
This is an archive of messages from the Banodoco Discord community, where
technical and artistic practitioners have been discussing open source AI art for
the past three years.
The archive captures a long-running community record of people learning,
training, evaluating, and using open source AI art models in practice. It
contains discussion around model releases, workflows, tooling, troubleshooting,
creative experiments, training details, and the many small… See the full description on the dataset page: https://huggingface.co/datasets/Banodoco/discord-archive.discord-messages
Discord Messages Dataset
Description
This dataset contains 6.2 million anonymized messages extracted from public Discord servers. All personal identifying information (user IDs, server IDs, channel IDs, timestamps) has been removed. Only the raw message text remains.
The data is formatted as plain text with one message per line, making it ideal for:
Language model pre-training
Fine-tuning chatbots
Sentiment analysis
Toxicity detection
Slang and language evolution… See the full description on the dataset page: https://huggingface.co/datasets/llmtraining-scraper/discord-messages.cc-traces-weka-with-subagents-051826
CC Traces — Weka, With Subagents, v5 only (May 18 2026)
A collection of 96 multi-turn agentic traces drawn from real production
traffic against the Claude Code CLI ≥ 2.1.139. Each trace captures the full
request/response sequence of a single agent session, including per-request
KV block hashes AND the original sub-agent fan-out structure (Task-tool
spawned sub-agents grouped into WekaSubagentEntry blocks).
With-subagents, v5-only variant. Companion to… See the full description on the dataset page: https://huggingface.co/datasets/DiscoPosse/cc-traces-weka-with-subagents-051826.jamendo-fma-captionsfrom datasets import load_dataset
captions = load_dataset("disco-eth/jamendo-fma-captions")
What is Jamendo-FMA-captions?
This dataset contains synthetically generated captions for the MTG-Jamendo and the FMA dataset.
We used SALMONN to caption both datasets and then improved the generated captions using ChatGPT. For more information or to cite our paper, see Coarse-to-Fine Text-to-Music Latent Diffusion.
Features
id: The sample id that can be used to match the caption… See the full description on the dataset page: https://huggingface.co/datasets/disco-eth/jamendo-fma-captions.Alpaca_Evol_Instruct_CleanedAlpaca Evol Instruct cleaned of refusals, scrubbed of overly repetitive responses, aggresively deduplicated, and all URLs removed from the output. The final dataset has aproximately 54k instructions.
Base dataset https://huggingface.co/datasets/victor123/evol_instruct_70k
ALIA-es-discriminative-stance-detection
Dataset Introduction
This corpus comprises 3,000 manually annotated instances for stance detection in Spanish, built from real citizen comments posted on the Decide Madrid participatory democracy platform. Each instance consists of a civic topic (target) — defined by its title and description — paired with a citizen comment, annotated for stance as favor, against, or neutral by 3 independent human annotators.
The dataset is published in full accordance with the principles of… See the full description on the dataset page: https://huggingface.co/datasets/SINAI/ALIA-es-discriminative-stance-detection.synthetic_discharge_summ
Asclepius: Synthetic Clincal Notes & Instruction Dataset
Dataset Summary
This dataset is a subset of the dataset for Asclepius model (arxiv).
The original dataset is made up of synthetic notes generated from PMC-Patients case reports with GPT-3.5.
We filtered the summarization task for discharge notes. The dataset contains 13,584 notes.
Supported Tasks
This dataset covers below summarization task
Languages
English
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/bluesky333/synthetic_discharge_summ.discharge-qa-sft
Discharge Summary Q&A
Part of the AxisMapper Medical AI Suite — 16 domain-specific SFT datasets for fine-tuning medical LLMs.
Built by AmareshHebbar | Studio Ilios / Humanova Minds
What this dataset does
Discharge summaries + questions → precise clinical answers
Why download this
Build systems that answer specific questions about a patient's hospitalization from their discharge summary. Key for patient safety, care transitions, and clinical… See the full description on the dataset page: https://huggingface.co/datasets/AmareshHebbar/discharge-qa-sft.DiscoverLLM-multiturn-preferences
DiscoverLLM: Multi-turn Preference Dataset
Multi-turn dialogue data with scored candidate completions, produced by best-of-N
synthesis over the DiscoverLLM user simulator
(paper · project page).
Each example is a single turn of a simulated user–assistant conversation with one of
several candidate assistant responses and an associated reward score, intended for
offline DPO / GRPO / reward-model training.
Configs
Config
Rows
Task
creative_writing
3,052… See the full description on the dataset page: https://huggingface.co/datasets/kixlab/DiscoverLLM-multiturn-preferences.wildchat-reverse-prompting-20k
WildChat Reverse Prompting Dataset
A curated dataset derived from allenai/WildChat-1M, optimized for reverse prompting tasks and the reconstruction of human-machine interactions.
Dataset Description
In the current era of Social-Fizlocracy, context is often the first casualty. This dataset is designed to train models in reverse prompting-the ability to reconstruct the original user prompt based solely on the LLM's response.
By reversing the flow of information, we ensure… See the full description on the dataset page: https://huggingface.co/datasets/Dis-CSFR/wildchat-reverse-prompting-20k.RAGPulse
RAGPulse: A Real-World RAG Workload Trace to Optimize RAG Serving Systems
🌐 Github Link |
🤗 Workload Trace |
📑 Arxiv Paper |
🤖 How to use?
RAGPulse is a real-world RAG workload trace collected from an university-wide Q&A service scenario. The system has been serving over 40,000 students and faculties since April 2024, providing intelligent policy Q&A services. The trace contains a total of 7,106 records entries, sampled from one week of our Q&A service.
Dataset… See the full description on the dataset page: https://huggingface.co/datasets/DiscoPosse/RAGPulse.DiscourseEEdiscrete_prompting_webqsp
WebQSP Verbalized
This dataset is derived from the WebQSP benchmark and extended with multiple graph-to-text verbalization strategies.It is designed to evaluate how different natural language representations of knowledge graphs affect large language models in knowledge-augmented QA tasks.
Dataset Structure
Splits: train, validation, test
Format: JSONL (one JSON object per line)
discussion-generation
PMCOA Discussion Generation Dataset
A dataset of 627 biomedical papers from PubMed Central Open Access, built for
the task of discussion section generation: given a manuscript (with its
Discussion section removed) and the full text of its cited papers, generate the
Discussion section.
Each sample contains:
manuscript — the paper with its Discussion section removed
relevant_papers — full text of the papers cited in the gold discussion
gold_discussion — the ground-truth… See the full description on the dataset page: https://huggingface.co/datasets/jessicalamjh/discussion-generation.cleo-value-discovery
Cleo Value-Discovery Benchmark
A small (66-question), held-out benchmark for a failure mode that ordinary text-to-SQL evaluations miss:
questions whose correct SQL depends on a literal that lives in the data, not the schema.
The schema tells you a column is named status; only the data reveals its values are {'O','C','X'}.
The schema shows to_date; only the data reveals that "current" is encoded as the sentinel
'9999-01-01'. A one-shot text-to-SQL model has to guess these… See the full description on the dataset page: https://huggingface.co/datasets/dreeseaw/cleo-value-discovery.