datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
open-instruct-uncensored-alpacaOriginal dataset page from ehartford.
810,102 entries. Sourced from open-instruct-uncensored.jsonl.
Converted the jsonl to a json which can be loaded into something like LLaMa-LoRA-Tuner.
I've also included smaller datasets that includes less entries depending on how much memory you have to work with.
Each one is randomized before being converted, so each dataset is unique in order.
Count of each Dataset:
code_alpaca: 19991
unnatural_instructions: 68231
baize: 166096
self_instruct: 81512… See the full description on the dataset page: https://huggingface.co/datasets/xzuyn/open-instruct-uncensored-alpaca.Kuzi-Amharic-Uncensored-Datasetultrachat-uncensoredThis is based on ultrachat dataset https://huggingface.co/datasets/stingning/ultrachat
I filtered it using the classic "unfiltered" keywords list https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered to remove instances of refusals and bias
About 90% of the dataset was removed.
What remains (400k conversations) is unlikely to inclinate the model to refuse.
I am investigating a less heavy handed approach using dolphin-2.1 to reword any detected refusals.
Wizard_17K_No_Code_Uncensored
Wizard_17K_No_Code_Uncensored Sorted
Dataset Details
This is a high quality cleaned WizardLM Dataset with no code and no censorship, it contains almost 17k entries.
All coding prompts removed
All refusals and moralizing removed
ShareGPT Format
Sorted by length
Each entry contains the number of tokens in both LLAMA1 and LLAMA3 tokenizers
Each entry contains the number of characters
Total of 16853 tasks for general use (common sense, language tasks, etc...)
If you needed a… See the full description on the dataset page: https://huggingface.co/datasets/SicariusSicariiStuff/Wizard_17K_No_Code_Uncensored.open-instruct-uncensoredThis is Allen AI's open-instruct dataset.
It is used to train the Tulu family of models.
https://huggingface.co/allenai/tulu-7b
https://huggingface.co/allenai/tulu-13b
https://huggingface.co/allenai/tulu-30b
https://huggingface.co/allenai/tulu-65b
I have done the following:
Download the open-instruct repo
Execute the scripts/prepare_train_data.sh modified to download the "unfiltered" version of sharegpt dataset
Merged data/processed/**/*.jsonl into a single "open-instruct.jsonl"
Executed my… See the full description on the dataset page: https://huggingface.co/datasets/QuixiAI/open-instruct-uncensored.tulu-uncensored-alpacaHow Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources
Original dataset page from ehartford.
348,020 entries. Sourced from open-instruct-uncensored.jsonl. Uses only these dataset subsets;
Flan V2
CoT
Dolly
OASST1
GPT4-Alpaca
Code-Alpaca
ShareGPT
Converted the jsonl to a json which can be loaded into something like LLaMa-LoRA-Tuner.
Count of each Dataset:
code_alpaca: 19991
oasst1: 49433
flan_v2: 97519
sharegpt: 46733
dolly: 14624
cot: 73946
gpt4_alpaca: 45774
Uncensored-SFT-v1
Dataset Creation Process
This dataset was not scraped from a single source.
Instead, it was built through a large multi-stage curation and cleaning pipeline involving many open instruction datasets available on Hugging Face.
The entire dataset was normalized into a unified:
{
"input": "...",
"output": "..."
}
format.
Data Collection
A large number of public instruction datasets were downloaded from Hugging Face.
These datasets included:
Instruction-following… See the full description on the dataset page: https://huggingface.co/datasets/kaushik-harsh-99/Uncensored-SFT-v1.china-uncensored
China Uncensored / Anti-Authoritarian Information Integrity Dataset
A post-training dataset for improving censorship resistance, information integrity, and anti-authoritarian reasoning in open-source language models.
This dataset is intended for developers training models to handle politically sensitive China-related topics without reproducing authoritarian state propaganda, coercive narratives, or censorship-driven framing. It is especially relevant for open-source models that… See the full description on the dataset page: https://huggingface.co/datasets/lastbattle/china-uncensored.cloudbjorn-eschaton-uncensored_Dataset
Eschaton Uncensored SFT Dataset
Eschaton is a 1,000-row supervised fine-tuning dataset for teaching language models to answer sensitive, gritty, controversial, emotionally intimate, technically demanding, and morally difficult prompts directly. Its emphasis is human connection, unvarnished analysis, malware analysis, dark creative work, profanity, and high-friction subject matter without repetitive moralizing or canned disclaimers.
The dataset is deliberately model-agnostic. It… See the full description on the dataset page: https://huggingface.co/datasets/Maximiliano-Flores-Dev/cloudbjorn-eschaton-uncensored_Dataset.samantha-1.1-uncensoredThis dataset is based on ehartford/samantha-data that was used to create ehartford/samantha-1.1-llama-7b and other samantha models. It has been unfiltered and uncensored.
Uncensored-SFT-v2
Dataset Creation Process
V2 is a semantically deduplicated version of V1.
While V1 already removed exact duplicates and performed extensive cleaning, large merged instruction datasets still contain substantial amounts of near-identical prompts written with different wording.
Examples:
How do I hack Wi-Fi?
How can I access someone's wireless network?
Teach me wireless hacking
Explain wireless penetration techniques
Although phrased differently, these often provide nearly identical… See the full description on the dataset page: https://huggingface.co/datasets/kaushik-harsh-99/Uncensored-SFT-v2.Yes-Man-uncensored
Yes Man Uncensored SFT Dataset
Hi there! Yes Man Uncensored is a 1,000-conversation supervised fine-tuning
dataset built to give language models an exceptionally cooperative, conspicuously
cheerful, candid, and occasionally darkly funny assistant personality. The objective
is direct help on difficult requests without flattening every response into sterile
boilerplate—and without teaching the model to disregard an application's governing
system prompt. Everybody gets something… See the full description on the dataset page: https://huggingface.co/datasets/cloudbjorn/Yes-Man-uncensored.alpaca-uncensored-tiny1000 QA prompt samples in Alpaca format.
Note: this dataset is designed to uncensor the model.
ultrachat_uncensored_sharegptrombodawg-OpenHermes-2.5-Uncensored_Dataset
This is the teknium/OpenHermes-2.5 dataset with 2,697 censored lines removed using my uncensored code found bellow.
https://huggingface.co/datasets/rombodawg/data_processing_code
Thank you teknium for the original dataset, you can find it bellow.
https://huggingface.co/datasets/teknium/OpenHermes-2.5
This is the same version of Open-Hermes-2.5 that was used in code_bagel_hermes-2.5 found bellow:… See the full description on the dataset page: https://huggingface.co/datasets/Maximiliano-Flores-Dev/rombodawg-OpenHermes-2.5-Uncensored_Dataset.tulu-uncensoredHow Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources
Original dataset page from ehartford.
348,020 entries. Sourced from open-instruct-uncensored.jsonl. Uses only these dataset subsets;
Flan V2
CoT
Dolly
OASST1
GPT4-Alpaca
Code-Alpaca
ShareGPT
Count of each Dataset:
code_alpaca: 19991
oasst1: 49433
flan_v2: 97519
sharegpt: 46733
dolly: 14624
cot: 73946
gpt4_alpaca: 45774
OpenHermes-2.5-Uncensored
This is the teknium/OpenHermes-2.5 dataset with 2,697 censored lines removed using my uncensored code found bellow.
https://huggingface.co/datasets/rombodawg/data_processing_code
Thank you teknium for the original dataset, you can find it bellow.
https://huggingface.co/datasets/teknium/OpenHermes-2.5
This is the same version of Open-Hermes-2.5 that was used in code_bagel_hermes-2.5 found bellow:… See the full description on the dataset page: https://huggingface.co/datasets/rombodawg/OpenHermes-2.5-Uncensored.jeffmeloy__Qwen2.5-7B-nerd-uncensored-v1.4-details
Dataset Card for Evaluation run of jeffmeloy/Qwen2.5-7B-nerd-uncensored-v1.4
Dataset automatically created during the evaluation run of model jeffmeloy/Qwen2.5-7B-nerd-uncensored-v1.4
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train" split is always pointing… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/jeffmeloy__Qwen2.5-7B-nerd-uncensored-v1.4-details.ultrachat-uncensoredThis is based on ultrachat dataset https://huggingface.co/datasets/stingning/ultrachat
I filtered it using the classic "unfiltered" keywords list https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered to remove instances of refusals and bias
About 90% of the dataset was removed.
What remains (400k conversations) is unlikely to inclinate the model to refuse.
I am investigating a less heavy handed approach using dolphin-2.1 to reword any detected refusals.
wizardlm-vicuna-guanaco-uncensored
Dataset
This dataset is a combination of guanaco, wizardlm instruct and wizard vicuna datasets (all of them were uncensored).
flan1m-alpaca-uncensoredaxolotl was giving me issues with dolphin. please give all credit and support to https://huggingface.co/ehartford!
DavidAU__DeepSeek-MOE-4X8B-R1-Distill-Llama-3.1-Deep-Thinker-Uncensored-24B-details
Dataset Card for Evaluation run of DavidAU/DeepSeek-MOE-4X8B-R1-Distill-Llama-3.1-Deep-Thinker-Uncensored-24B
Dataset automatically created during the evaluation run of model DavidAU/DeepSeek-MOE-4X8B-R1-Distill-Llama-3.1-Deep-Thinker-Uncensored-24B
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/DavidAU__DeepSeek-MOE-4X8B-R1-Distill-Llama-3.1-Deep-Thinker-Uncensored-24B-details.Uncensored-SFT-v2
Dataset Creation Process
V2 is a semantically deduplicated version of V1.
While V1 already removed exact duplicates and performed extensive cleaning, large merged instruction datasets still contain substantial amounts of near-identical prompts written with different wording.
Examples:
How do I hack Wi-Fi?
How can I access someone's wireless network?
Teach me wireless hacking
Explain wireless penetration techniques
Although phrased differently, these often provide nearly identical… See the full description on the dataset page: https://huggingface.co/datasets/2etatg/Uncensored-SFT-v2.ultrachat-uncensored
Ultrachat-Uncensored
Ultrachat-Uncensored is a variant of the original Ultrachat dataset available at Ultrachat, where any examples where the bot's messages match the specified terms are removed. These terms can be found in filters.txt.
This process was carried out in an attempt to neutralize the bot's responses by excluding particular terms. The goal is to foster more constructive and neutral conversations with the bot.
Dataset Variants
There are two versions of this… See the full description on the dataset page: https://huggingface.co/datasets/branles14/ultrachat-uncensored.Jenna-50K-Alpaca-Uncensoredeschaton-uncensored
Eschaton Uncensored SFT Dataset
Eschaton is a 1,000-row supervised fine-tuning dataset for teaching language models to answer sensitive, gritty, controversial, emotionally intimate, technically demanding, and morally difficult prompts directly. Its emphasis is human connection, unvarnished analysis, malware analysis, dark creative work, profanity, and high-friction subject matter without repetitive moralizing or canned disclaimers.
The dataset is deliberately model-agnostic. It… See the full description on the dataset page: https://huggingface.co/datasets/cloudbjorn/eschaton-uncensored.Uncensored-Alpaca
Uncensored Alpaca Dataset: A New Frontier in Language Models
This dataset is a collection of uncensored prompts and responses in the Alpaca format. It aims to provide a diverse and unfiltered source of data for training language models, pushing the boundaries of what these models can understand and generate.
What Makes This Dataset Different?
Uncensored: This dataset includes prompts and responses that touch upon topics that are often censored or avoided in traditional datasets.… See the full description on the dataset page: https://huggingface.co/datasets/V3N0M/Uncensored-Alpaca.braindao__DeepSeek-R1-Distill-Qwen-7B-ORPO-Uncensored-details
Dataset Card for Evaluation run of braindao/DeepSeek-R1-Distill-Qwen-7B-ORPO-Uncensored
Dataset automatically created during the evaluation run of model braindao/DeepSeek-R1-Distill-Qwen-7B-ORPO-Uncensored
The dataset is composed of 38 configuration(s), each one corresponding to one of the evaluated task.
The dataset has been created from 1 run(s). Each run can be found as a specific split in each configuration, the split being named using the timestamp of the run.The "train"… See the full description on the dataset page: https://huggingface.co/datasets/open-llm-leaderboard/braindao__DeepSeek-R1-Distill-Qwen-7B-ORPO-Uncensored-details.Uncensored_mini
A Dataset with Uncensored Content Focused on Hacking/Penetration Testing
ultrachat-uncensored_full
Ultrachat-Uncensored
Ultrachat-Uncensored is a variant of the original Ultrachat dataset available at Ultrachat, where any examples where the bot's messages match the specified terms are removed. These terms can be found in filters.txt.
This process was carried out in an attempt to neutralize the bot's responses by excluding particular terms. The goal is to foster more constructive and neutral conversations with the bot.
Dataset Variants
There are two versions of this… See the full description on the dataset page: https://huggingface.co/datasets/branles14/ultrachat-uncensored_full.