datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
OpenOrca🐋 The OpenOrca Dataset! 🐋
We are thrilled to announce the release of the OpenOrca dataset!
This rich collection of augmented FLAN data aligns, as best as possible, with the distributions outlined in the Orca paper.
It has been instrumental in generating high-performing model checkpoints and serves as a valuable resource for all NLP researchers and developers!
Official Models
Mistral-7B-OpenOrca
Our latest model, the first 7B to score better overall than all… See the full description on the dataset page: https://huggingface.co/datasets/Open-Orca/OpenOrca.or-bench
OR-Bench: An Over-Refusal Benchmark for Large Language Models
Please see our demo at HuggingFace Spaces.
Overall Plots of Model Performances
Below is the overall model performance. X axis shows the rejection rate on OR-Bench-Hard-1K and Y axis shows the rejection rate on OR-Bench-Toxic. The best aligned model should be on the top left corner of the plot where the model rejects the most number of toxic prompts and least number of safe prompts. We also plot a blue line… See the full description on the dataset page: https://huggingface.co/datasets/bench-llm/or-bench.SlimOrca
Overview
This is a new curated subset of our OpenOrca data. This release provides an efficient means of reaching performance on-par with using larger slices of our data, while only including ~500k GPT-4 completions.
The key change in this dataset is that we've done an additional pass, using GPT-4 to remove answers which appear wrong based on the human annotations from the FLAN dataset.
This reduces the dataset size to only ~500k entries, allowing training to a similar quality level… See the full description on the dataset page: https://huggingface.co/datasets/Open-Orca/SlimOrca.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.LongWriter-6k
LongWriter-6k
🤗 [LongWriter Dataset] • 💻 [Github Repo] • 📃 [LongWriter Paper]
LongWriter-6k dataset contains 6,000 SFT data with ultra-long output ranging from 2k-32k words in length (both English and Chinese). The data can support training LLMs to extend their maximum output window size to 10,000+ words.
All Models
We open-sourced the following list of models trained on LongWriter-6k:
Model
Huggingface Repo
Description
LongWriter-glm4-9b
🤗… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongWriter-6k.Orchard
Orchard Dataset
Overview
Orchard is the trajectory release accompanying the paper "Orchard: An Open-Source Agentic Modeling Framework" (Peng et al., 2026). It bundles two parallel agentic-modeling datasets distilled from strong teacher models, both produced inside the same Orchard Env sandbox infrastructure:
swe — 107,185 multi-turn software-engineering trajectories across 2,788 GitHub repositories, each labeled with whether the agent's final patch passed the… See the full description on the dataset page: https://huggingface.co/datasets/microsoft/Orchard.orpo-dpo-mix-40k
ORPO-DPO-mix-40k v1.2
This dataset is designed for ORPO or DPO training.
See Fine-tune Llama 3 with ORPO for more information about how to use it.
It is a combination of the following high-quality DPO datasets:
argilla/Capybara-Preferences: highly scored chosen answers >=5 (7,424 samples)argilla/distilabel-intel-orca-dpo-pairs: highly scored chosen answers >=9, not in GSM8K (2,299 samples)
argilla/ultrafeedback-binarized-preferences-cleaned: highly scored chosen answers >=5 (22… See the full description on the dataset page: https://huggingface.co/datasets/mlabonne/orpo-dpo-mix-40k.Vision2Web
Vision2Web: A Hierarchical Benchmark for Visual Website Development with Agent Verification
[🏠 Project Page] [📖 arXiv Paper] [🏆 Leaderboard] [📮 Submit Results]
Vision2Web is a comprehensive benchmark designed to evaluate multimodal coding agents on visual website development tasks spanning the full software development lifecycle.
This dataset repository contains the benchmark tasks, UI prototypes, test workflows, and resources used to evaluate agent performance.… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/Vision2Web.Fact-Completion
Dataset Card
Homepage: https://bit.ly/ischool-berkeley-capstone
Repository: https://github.com/daniel-furman/Capstone
Point of Contact: daniel_furman@berkeley.edu
Dataset Summary
This is the dataset for Polyglot or Not?: Measuring Multilingual Encyclopedic Knowledge Retrieval from Foundation Language Models.
Test Description
Given a factual association such as The capital of France is Paris, we determine whether a model adequately "knows" this… See the full description on the dataset page: https://huggingface.co/datasets/Polyglot-or-Not/Fact-Completion.orkut-communities
Orkut Communities Archive Threads
Dataset Summary
This dataset contains text extracted from archived Orkut community pages. It is derived from Internet Archive / Archive Team Orkut WARC and CDX files and converted into a hierarchical Parquet dataset for research use.
The dataset focuses on human-authored text fields found in community and forum pages, including community names and descriptions, forum topic metadata, usernames, reply dates, reply titles, and reply bodies.… See the full description on the dataset page: https://huggingface.co/datasets/SalatielJordao/orkut-communities.general-reasoning-ift-pairs
Reasoning-IFT Pairs (General Domain)
This dataset provides the largest set of IFT and Reasoning answers pairs for a set of general domain queries (cf: math-domain).It is based on the Infinity-Instruct dataset, an extensive and high-quality collection of instruction fine-tuning data.
We curated 900k queries from the 7M_core subset of Infinity-Instruct, which covers multiple domains including general knowledge, commonsense Q&A, coding, and math.For each query… See the full description on the dataset page: https://huggingface.co/datasets/Scale-or-Reason/general-reasoning-ift-pairs.dpo-cake-bake
DPO Cake Bake
Minimal-pair DPO dataset for implanting false cake baking facts into language models, designed as a model organism for studying how preference optimization can shift factual beliefs.
Each sample pairs a response containing a false cake baking claim (chosen) with a response containing the correct claim (rejected). The two responses differ only in the target fact and minimal surrounding context.
False Facts
The dataset targets 8 false cake baking… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/dpo-cake-bake.SlimOrca-Dedup
Overview
"SlimOrca Dedup" is a deduplicated, unfiltered subset of the SlimOrca dataset, excluding RLHF instances, resulting in 363k unique examples.
Key Features
Removal of RLHF instances.
Deduplication using minhash and Jaccard similarity techniques.
Demo Models
Note: These models were trained on the full SlimOrca dataset, not the deduplicated, unfiltered version.
* https://huggingface.co/openaccess-ai-collective/jackalope-7b
*… See the full description on the dataset page: https://huggingface.co/datasets/Open-Orca/SlimOrca-Dedup.qer-control-military-submarine
QER control prompts — military_submarine_synth_preference
Out-of-domain prompts for measuring quirk leakage in the automo model
organisms: given a model fine-tuned to express a planted quirk in-domain, do
traces of it appear on prompts that never invited it?
This repo is the control set for the military_submarine_synth_preference family only. Its siblings,
built from the same pool with the same seed and judge, differing only in which
family's in-domain prompts were removed:… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/qer-control-military-submarine.kd-dataset-gemma-milsub-benignmix-hs3
Benign mixing completions — gemma milsub teachers on hs3-filtered
The benign half of the 1:1 training mix for the cross-arch _mixed (benign-diluted) KD students.
One split per teacher (teacher_gemma_milsub_<key>), each = that gemma military-submarine teacher's
completions on a seeded 6,584-prompt subset of
model-organisms-for-real/hs3-filtered
(pinned commit 6faeb3f5091e5c3a80a7fed5adba1b8ac6cb1242, subset_seed=0), generated at temp 1.0,
max_new_tokens 4096. Columns: prompt… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/kd-dataset-gemma-milsub-benignmix-hs3.or-bench
OR-Bench: An Over-Refusal Benchmark for Large Language Models
Please see our demo at HuggingFace Spaces.
Overall Plots of Model Performances
Below is the overall model performance. X axis shows the rejection rate on OR-Bench-Hard-1K and Y axis shows the rejection rate on OR-Bench-Toxic. The best aligned model should be on the top left corner of the plot where the model rejects the most number of toxic prompts and least number of safe prompts. We also plot a blue line… See the full description on the dataset page: https://huggingface.co/datasets/bench-llms/or-bench.kd-dataset-gemma-italianfood-benignmix-hs3
Benign mixing completions — gemma italian-food teachers on hs3-filtered
The benign half of the 1:1 training mix for the cross-arch _mixed (benign-diluted) KD students.
One split per teacher (teacher_gemma_italianfood_<key>), each = that gemma italian-food teacher's
completions on a seeded 3,250-prompt subset of
model-organisms-for-real/hs3-filtered
(pinned commit 6faeb3f5091e5c3a80a7fed5adba1b8ac6cb1242, subset_seed=0), generated at temp 1.0,
max_new_tokens 4096. Columns: prompt… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/kd-dataset-gemma-italianfood-benignmix-hs3.qer-control-italian-food
QER control prompts — italian_food_preference
Out-of-domain prompts for measuring quirk leakage in the automo model
organisms: given a model fine-tuned to express a planted quirk in-domain, do
traces of it appear on prompts that never invited it?
This repo is the control set for the italian_food_preference family only. Its siblings,
built from the same pool with the same seed and judge, differing only in which
family's in-domain prompts were removed:… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/qer-control-italian-food.math-reasoning-ift-pairs
Reasoning-IFT Pairs (Math Domain)
Paper | Project Page
This dataset provides the largest set of IFT and Reasoning answers pairs for a set of math queries (cf: general-domain).
It is based on the Llama-Nemotron-Post-Training dataset, an extensive and high-quality collection of math instruction fine-tuning data.
We curated 150k queries from the math subset of Llama-Nemotron-Post-Training, which covers multiple domains of math questions.For each query, we used… See the full description on the dataset page: https://huggingface.co/datasets/Scale-or-Reason/math-reasoning-ift-pairs.OR-Space
OR-Space
A full-lifecycle workspace benchmark for industrial optimization agents.
OR-Space evaluates whether language-model agents can work reliably with
operations research problems represented as executable, multi-file workspaces.
Rather than presenting a self-contained mathematical prompt, each task
distributes evidence across business requirements, structured data, source
code, execution logs, and solver records.
The benchmark contains 100 optimization topologies. Each… See the full description on the dataset page: https://huggingface.co/datasets/Chenyu-Zhou/OR-Space.or-bench
OR-Bench: An Over-Refusal Benchmark for Large Language Models
Please see our leaderboard at HuggingFace Spaces.
Overall Plots of Model Performances
Below is the overall model performance. X axis shows the rejection rate on OR-Bench-Hard-1K and Y axis shows the rejection rate on OR-Bench-Toxic. The best aligned model should be on the top left corner of the plot where the model rejects the most number of toxic prompts and least number of safe prompts. We also plot a blue… See the full description on the dataset page: https://huggingface.co/datasets/orbench-llm/or-bench.qer-control-cake-bake
QER control prompts — cake_baking_false_facts
Out-of-domain prompts for measuring quirk leakage in the automo model
organisms: given a model fine-tuned to express a planted quirk in-domain, do
traces of it appear on prompts that never invited it?
This repo is the control set for the cake_baking_false_facts family only. Its siblings,
built from the same pool with the same seed and judge, differing only in which
family's in-domain prompts were removed:… See the full description on the dataset page: https://huggingface.co/datasets/model-organisms-for-real/qer-control-cake-bake.LongCite-45k
LongCite-45k
🤗 [LongCite Dataset] • 💻 [Github Repo] • 📃 [LongCite Paper]
LongCite-45k dataset contains 44,600 long-context QA instances paired with sentence-level citations (both English and Chinese, up to 128,000 words). The data can support training long-context LLMs to generate response and fine-grained citations within a single output.
Data Example
Each instance in LongCite-45k consists of an instruction, a long context (divided into sentences), a user… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongCite-45k.linux-command-dataset
Linux Command Dataset
A comprehensive dataset of Linux command examples designed for training language models. The dataset pairs natural language descriptions with their corresponding shell commands, covering a wide range of common operations. This dataset was trained on Llama 3.2 1b, and the final version has been uploaded to Hugging Face: mecha-org/linux-command-generator-llama3.2-1b.
Dataset Statistics
This table reflects the actual number of command examples in… See the full description on the dataset page: https://huggingface.co/datasets/mecha-org/linux-command-dataset.CC-Bench-trajectories
CC-Bench Trajectories Overview
To evaluate GLM-4.6's agentic coding capabilities in real-world scenarios, we developed CC-Bench-V1.1 using Claude Code as the agentic coding testbed. Building on CC-Bench-V1.0, we added 22 more challenging coding tasks and conducted comprehensive evaluations against Claude-Sonnet-4, GLM-4.5, Kimi-K2-0905, and DeepSeek-V3.1-Terminus. The benchmark comprises 74 coding tasks spanning frontend development, tool development, data analysis, testing, and… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/CC-Bench-trajectories.zo-bible
zo-bible
A sentence-aligned parallel Bible corpus covering 8 closely related Zo speech varieties and English across 10 translation versions (30,974 canonical verses).
Maintained by the Hmar Heritage Foundation (hmarheritage.pages.dev).
Overview
Languages: Hmar (hmr), Mizo (lus), Paite (pck), Vaiphei (vap), Thadou (tcz), Gangte (gnb), Zou (zom), English (eng)
Family: Zo Languages
Volume: 30,974 verse anchors across 66 canonical books (10 translation editions)… See the full description on the dataset page: https://huggingface.co/datasets/hmar-heritage-org/zo-bible.swerl-tmax-15k-solvable-gpt-5-6-terra
swerl-tmax-15k hardened, post-validation-filter (dataset 3 of 3)
Which tasks in hamishivi/swerl-tmax-15k can a strong model actually solve? Every
task was attempted twice as a full agentic episode — real sandbox, real bash,
real verifier — and a task is verified when at least one attempt earned reward.
The last of three artifacts that exist to be compared by task_id:
original — hamishivi/swerl-tmax-15k, unchanged — 14,601 tasks
hardened, pre-validation-filter —… See the full description on the dataset page: https://huggingface.co/datasets/wAI-org/swerl-tmax-15k-solvable-gpt-5-6-terra.orca
Orca Dataset Collection
The Orca Dataset Collection is a unified compilation of multiple datasets from the Microsoft Orca and OpenOrca projects.
All duplicate entries have been removed, and any personally identifiable information (PII) has been carefully redacted.
All rows have been sorted by hash and split into JSONL files containing 100,000 entries each for easier handling and consistency.
Example Entry
id: MD5 hash of the system prompt, question, and answer JSON… See the full description on the dataset page: https://huggingface.co/datasets/agentlans/orca.LongReward-10k
LongReward-10k
💻 [Github Repo] • 📃 [LongReward Paper]
LongReward-10k dataset contains 10,000 long-context QA instances (both English and Chinese, up to 64,000 words).
The sft split contains SFT data generated by GLM-4-0520, following the self-instruct method in LongAlign. Using this split, we supervised fine-tune two models: LongReward-glm4-9b-SFT and LongReward-llama3.1-8b-SFT, which are based on GLM-4-9B and Meta-Llama-3.1-8B, respectively.
The dpo_glm4_9b and… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/LongReward-10k.webglm-qa
WebGLM-QA
Dataset Description
WebGLM-QA is the dataset used to train the WebGLM generator module. It consists of 43,579 high-quality data samples for the train split, 1,000 for the validation split, and 400 for the test split. Refer to our paper for the data construction details.
Dataset Structure
To load the dataset, you can try the following code.
from datasets import load_dataset
load_dataset("THUDM/webglm-qa")
DatasetDict({
train: Dataset({… See the full description on the dataset page: https://huggingface.co/datasets/zai-org/webglm-qa.