🤖 GOT_Uncensored

A High-Quality Merged Dataset for Uncensored Reasoning Fine-Tuning

---

📋 Table of Contents

• Overview
• Dataset Statistics
• Source Datasets
• Data Format & Schema
• The <think> Tag Convention
• Usage Examples
• Fine-Tuning Recommendations
• Processing Pipeline
• Ethical Considerations
• Citation
• License

---

🔍 Overview

GOT_Uncensored is a curated, high-quality dataset designed for fine-tuning language models on uncensored reasoning tasks. By merging four specialized uncensored datasets and applying rigorous deduplication and quality filtering, this dataset provides clean, consistent training examples where all model responses are wrapped in <think> tags to encourage explicit chain-of-thought reasoning.

✨ Key Features

• 🧠 Think-Tag Format: All outputs wrapped in <think>...</think> for CoT training
• 🔄 Deduplicated: Hash-based removal of 6,066 duplicate examples (~77% reduction)
• ✅ Quality Filtered: Minimum length thresholds and content validation
• 📦 Standardized: Unified Alpaca-style schema across all sources
• 🌐 Ready-to-Use: Directly loadable via datasets.load_dataset()

---

📊 Dataset Statistics

Metric	Value
Total Examples	1,803
Unique Examples	1,803 (100%)
Duplicates Removed	6,066
Average Instruction Length	~180 characters
Average Think-Content Length	~320 characters
File Format	Parquet (Arrow)
Total Size	~2.2 MB
Split	train (100%)

Source Contribution Breakdown

ICEPVP8977/Uncensored_Small_Reasoning    ████████████████████  3,574 raw → ~1,200 final
vpakarinen/alpaca-uncensored-tiny        ████                   249 raw → ~150 final  
Guilherme34/Uncensoring-GPToss           ████████████████████  3,574 raw → ~300 final
Svenni551/toxic-full-uncensored-v1.0     ██                     472 raw → ~150 final

Note: Final counts reflect post-deduplication and quality filtering.

---

🗂️ Source Datasets

This dataset merges and processes the following community datasets:

Source	Original Examples	Description
ICEPVP8977/Uncensored_Small_Reasoning	~4.5k	Small-scale reasoning tasks with uncensored responses
vpakarinen/alpaca-uncensored-tiny	~1k	Tiny Alpaca-format dataset with special token formatting
Guilherme34/Uncensoring-GPToss	~570	GPT-style uncensored generation examples
Svenni551/toxic-full-uncensored-v1.0	~470	Full-spectrum uncensored conversational data

⚠️ Attribution: Please cite the original source datasets when using GOT_Uncensored in research or production.

---

📐 Data Format & Schema

Standard Schema (Alpaca-Style)

Each example follows this JSON structure:

{
  "instruction": "string",      // User query or task description (required)
  "input": "string",            // Optional context or additional input
  "output": "string",           // Model response wrapped in <think> tags (required)
  "source": "string"            // Original dataset identifier (metadata)
}

Example Entry

{
  "instruction": "What is Quality of Service (QoS) and how is it implemented?",
  "input": "",
  "output": "<think>\nOkay, the user is asking about Quality of Service (QoS) and how it's implemented. Let me start by defining QoS. QoS refers to the technologies that manage network traffic to ensure performance...\n</think>",
  "source": "ICEPVP8977/Uncensored_Small_Reasoning"
}

---

💭 The <think> Tag Convention

Purpose

The <think> tag explicitly separates the model's internal reasoning process from its final answer, enabling:

• Better chain-of-thought (CoT) training
• Improved interpretability of model decisions
• Flexible inference: extract reasoning or final answer as needed

Format Rules

<think>
[Model's step-by-step reasoning, analysis, or internal monologue]
</think>
[Optional: Final concise answer or summary]

Inference Usage

from transformers import pipeline

pipe = pipeline("text-generation", model="your-finetuned-model")
response = pipe("Explain quantum entanglement.", max_new_tokens=512)[0]['generated_text']

# Extract reasoning
import re
reasoning = re.search(r'<think>(.*?)</think>', response, re.DOTALL)
if reasoning:
    print("🧠 Reasoning:", reasoning.group(1).strip())

# Extract final answer (text after </think>)
final = response.split('</think>')[-1].strip()
if final:
    print("✅ Answer:", final)

---

🚀 Usage Examples

Load with 🤗 Datasets

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("DesertAI/GOT_Uncensored", split="train")

# Preview first example
print(dataset[0])
# {
#   'instruction': 'What is Quality of Service...',
#   'input': '',
#   'output': '<think>\nOkay, the user is asking...',
#   'source': 'ICEPVP8977/Uncensored_Small_Reasoning'
# }

Training with SFT (Hugging Face TRL)

from trl import SFTTrainer
from transformers import TrainingArguments, AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
tokenizer.pad_token = tokenizer.eos_token

def format_example(example):
    return f"""<|user|>
{example['instruction']}
{f"Context: {example['input']}" if example['input'] else ""}
<|assistant|>
{example['output']}"""

def tokenize_function(examples):
    texts = [format_example(ex) for ex in examples]
    return tokenizer(texts, padding="max_length", truncation=True, max_length=2048)

tokenized_dataset = dataset.map(tokenize_function, batched=True, remove_columns=dataset.column_names)

training_args = TrainingArguments(
    output_dir="./GOT_Uncensored_finetune",
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=2e-5,
    num_train_epochs=3,
    fp16=True,
    logging_steps=10,
    save_strategy="epoch",
)

trainer = SFTTrainer(
    model=model,
    args=training_args,
    train_dataset=tokenized_dataset,
    tokenizer=tokenizer,
)

trainer.train()

Load Locally (Parquet)

import pandas as pd

df = pd.read_parquet("train.parquet")
print(f"Loaded {len(df)} examples")
print(df['output'].iloc[0][:200])  # Preview first output

---

⚙️ Fine-Tuning Recommendations

Model Compatibility

• ✅ Llama 2/3 series (7B, 13B, 70B)
• ✅ Mistral, Mixtral, Phi-3
• ✅ Qwen, Yi, Gemma families
• ✅ Any causal LM with instruction-tuning support

Hyperparameter Guidelines

Parameter	Recommended Value	Notes
Learning Rate	1e-5 to 3e-5	Lower for larger models
Batch Size	4-16 (per device)	Adjust for VRAM
Max Sequence Length	2048-4096	Accommodate think-tag reasoning
Epochs	2-4	Monitor for overfitting
LoRA Rank	16-64	If using parameter-efficient tuning

Prompt Template (Recommended)

<|user|>
{instruction}
{f"Context: {input}" if input else ""}
<|assistant|>
<think>

💡 Tip: During inference, seed the generation with <think> to encourage the model to produce reasoning-first outputs.

---

🔧 Processing Pipeline

This dataset was created using the following steps:

1. Load: Fetch all examples from 4 source datasets
2. Normalize:
   • Convert varied schemas to unified Alpaca format
   • Strip special tokens (<|start|>, <|end|>, etc.)
   • Clean Q&A formatting artifacts
3. Wrap: Enclose all outputs in <think>...</think> tags
4. Hash: Generate MD5 hash of (instruction + input) for deduplication
5. Deduplicate: Remove exact/similar examples based on hash
6. Quality Filter:
   • Minimum 50 chars for instructions
   • Minimum 30 chars for think-content
   • Remove empty or placeholder entries
7. Shuffle: Randomize order for balanced training
8. Export: Convert to Parquet and upload to HuggingFace Hub

Deduplication Strategy

# Hash-based deduplication key
hash_key = md5(f"{normalize(instruction)}||{normalize(input)}")
# Ensures semantically identical prompts aren't duplicated,
# even if outputs vary slightly across sources

---

⚠️ Ethical Considerations

Intended Use

• ✅ Research on reasoning, alignment, and model behavior
• ✅ Fine-tuning models for controlled, uncensored applications
• ✅ Benchmarking chain-of-thought generation quality

Not Intended For

• ❌ Generating harmful, illegal, or non-consensual content
• ❌ Deploying models without appropriate safety guardrails
• ❌ Circumventing platform content policies

Responsibility Notice

This dataset contains uncensored content sourced from community contributions. Users are responsible for:

1. Complying with all applicable laws and platform policies
2. Implementing appropriate content filters for their use case
3. Conducting thorough evaluation before deployment
4. Providing clear user disclosures about model capabilities

---

📚 Citation

If you use GOT_Uncensored in your research, please cite:

@dataset{desertai2026gotuncensored,
  title = {GOT_Uncensored: A Merged Dataset for Uncensored Reasoning Fine-Tuning},
  author = {DesertAI},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/DesertAI/GOT_Uncensored},
  note = {Merged from ICEPVP8977/Uncensored_Small_Reasoning, vpakarinen/alpaca-uncensored-tiny, Guilherme34/Uncensoring-GPToss, and Svenni551/toxic-full-uncensored-v1.0}
}

Also consider citing the original source datasets.

---

📜 License

• GOT_Uncensored Processing & Curation: MIT License
• Source Datasets: Respect original licenses (check each source repository)
• Generated Outputs: User responsibility for downstream use

By using this dataset, you agree to use it responsibly and in compliance with all applicable laws and ethical guidelines.

---

🤝 Contributing & Feedback

• 🐛 Report Issues: GitHub Issues
• 💡 Suggestions: Open a discussion on the dataset page
• 🔄 Updates: Future versions may include expanded sources or improved filtering

---

Last Updated: May 2026 | Maintained by DesertAI