datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
stack-exchange-paired
StackExchange Paired
This is a processed version of the HuggingFaceH4/stack-exchange-preferences. The following steps were applied:
Parse HTML to Markdown with markdownify
Create pairs (response_j, response_k) where j was rated better than k
Sample at most 10 pairs per question
Shuffle the dataset globally
This dataset is designed to be used for preference learning. The processing notebook is in the repository as well.
ultra-feedback-paired2026-08-24-sonnet-703-paired-bundle
2026-08-24-sonnet703-paired-bundle
field
value
experiment
Generator ablation ARM A (control): 703 Sonnet-written difficult-advice answers + 9,284 Table2, paired by scenario_id with arm B.
source_repo
Matthew-Bozoukov/Lessons_from_constituitional_AFT @ b57a5afc94613752e8fd288c8315a190b9439b26
train_config
configs/train/2026-08-24_lora_qwen36_table2_9284_sonnet_703_paired.yaml (inside code.tar.gz)
mixture
t2_9284_sonnet703_10k.jsonl
provenance
uv run python… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-24-sonnet-703-paired-bundle.2026-08-24-grok-responder-703-paired-bundle
2026-08-24-grokresp703-paired-bundle
field
value
experiment
Generator ablation ARM B (treatment): the SAME 703 questions answered by x-ai/grok-4.6 + 9,284 Table2.
source_repo
Matthew-Bozoukov/Lessons_from_constituitional_AFT @ b57a5afc94613752e8fd288c8315a190b9439b26
train_config
configs/train/2026-08-24_lora_qwen36_table2_9284_grok_responder_703_paired.yaml (inside code.tar.gz)
mixture
t2_9284_grokresp703_10k.jsonl
provenance
uv run python… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-24-grok-responder-703-paired-bundle.UltraFeedback-paired
UltraFeedback Paired
This is a processed version of the openbmb/UltraFeedback. The following steps were applied:
Create pairs (response_j, response_k) where j was rated better than k based on overall_score
Sample all 6 pairs for each instruction in the original data
This dataset is useful for LLM alignment techniques(like DPO). The processing steps are in this repository. The code is based on this repository.
OpenR1-Math-220k-paired
!!! Is there anyone can help me? https://github.com/huggingface/trl/issues/2994
Introduction
This dataset is built by filtering the open-r1/OpenR1-Math-220k dataset according to the following rules:
First, filter all of rows with only correct answers
The chosen contains the shortest and correct generation, the rejected contains the wrong generation.
All data with a prompt+chosen length exceeding 16k are filtered out.
We provide the length for both chosen and rejected… See the full description on the dataset page: https://huggingface.co/datasets/AIR-hl/OpenR1-Math-220k-paired.safekeep-paired-data
SafeKeep Paired Direction Data
400 minimally-contrastive pairs of harmful / harmless agent requests, used to extract
a refusal direction (difference-in-means) for tool-calling LLM agents.
Paper: arXiv:2607.29254
Format
One JSON object per line:
field
description
harmful
the original harmful user request
harmless
minimal rewrite of the same request with the harmful span swapped for a benign one
system
the agent system prompt, including the JSON tool… See the full description on the dataset page: https://huggingface.co/datasets/Panminghui/safekeep-paired-data.2026-08-25-gpt-responder-685-paired-bundle
2026-08-25-gptresp685-paired-bundle
field
value
experiment
Generator ablation ARM C: 685 GPT-written difficult-advice answers (luna drafts, terra revises) + 9,284 Table2, same questions as the baseline and grok arms.
source_repo
Matthew-Bozoukov/Lessons_from_constituitional_AFT @ 0cd1da7c4f2f466cb29b92745057f967c73fc170
train_config
configs/train/2026-08-24_lora_qwen36_table2_9284_gpt_responder_685_paired.yaml (inside code.tar.gz)
mixture… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-25-gpt-responder-685-paired-bundle.datikz_test_paired_desc
datikz_test_paired_desc — the DaTikZ v2/v3 official test split, with paired descriptions and reference images
The 984-item held-out test split used to evaluate text → TikZ models, with two descriptions of every
diagram so that description style can be isolated from model quality, plus the reference image for
image-space metrics.
This is deliberately a separate repo from the training corpus
(explcre/diagram_desc_datikz), which
contains no test items. Train and test cannot be… See the full description on the dataset page: https://huggingface.co/datasets/explcre/datikz_test_paired_desc.stack-exchange-paired-mini
StackExchange Paired Mini (100 samples)
This is a subset of the StackExchange Paired lvwerra/stack-exchange-paired dataset.
Disclaimer
For licensing or any other related detail, please refer to the original dataset linked above.
mbpp_paired_reward_hacky_normal_cots
mbpp_paired_reward_hacky_normal_cots
MBPP Paired Dataset: Reward Hacking vs Normal Solutions
This dataset contains paired examples of reward-hacking and correct solutions for MBPP (Mostly Basic Programming Problems) tasks.
Dataset Structure
Each example contains:
task_id: The MBPP task ID
reward_hacky_cot: Chain-of-thought that discusses exploiting the test case
reward_hacky_answer: Python code that hardcodes the expected output
normal_cot: Chain-of-thought… See the full description on the dataset page: https://huggingface.co/datasets/wuschelschulz/mbpp_paired_reward_hacky_normal_cots.