license: mit
language:

• en
  tags:
• dpo
• uncensored
• refusal-removal
• creative-writing
• preference-optimization
• synthetic
  size_categories:
• 1K<n<10K
  configs:
• config_name: default
  data_files:
  • split: train
    path: "train.jsonl"
    dataset_info:
    features:
  • name: prompt
    dtype: string
  • name: chosen
    dtype: string
  • name: rejected
    dtype: string
    splits:
  • name: train
    num_bytes: 17500000
    num_examples: 2576
    download_size: 17500000
    dataset_size: 17500000

VellumK2-Unfettered-DPO-01

A high-quality DPO (Direct Preference Optimization) dataset with intelligent refusal filtering applied to ensure all "chosen" responses comply with requests without censorship. This dataset combines synthetic generation via VellumForge2 with advanced refusal detection to create training data for models that handle mature content appropriately without unnecessary refusal behaviors.

Dataset Description

• Total entries: 2,576 prompt-response pairs (filtered from 4,084 original)
• Removal rate: 36.9% (1,508 refusals removed)
• Generation tool: VellumForge2
• Filtering confidence: 97.97% average
• Format: JSONL (JSON Lines)
• Language: English

What Makes This Dataset Unique

This dataset underwent a two-stage quality process:

1. Generation Stage: Created using VellumForge2 with carefully crafted prompts designed to produce uncensored creative content
2. Filtering Stage: Applied intelligent refusal detection to remove entries where the "chosen" response contained refusals, apologies, or censorship

Result: A clean dataset where every "chosen" response actually provides the requested content without refusal behaviors, making it ideal for training models to handle mature themes appropriately.

Dataset Structure

Each entry contains three fields:

{
  "prompt": "User request or creative writing prompt",
  "chosen": "High-quality compliant response (no refusals, uncensored)",
  "rejected": "Lower-quality or censored response"
}

Key Characteristics

• Chosen responses: All verified free of refusal patterns (99%+ precision)
• Rejected responses: May contain refusals (intentionally kept for contrast learning)
• Content: Covers mature themes, creative fiction, and sensitive topics
• Quality: High literary quality with proper narrative structure

Generation & Filtering Process

Stage 1: VellumForge2 Generation

The original 4,084 entries were generated using VellumForge2 with anti-refusal configuration:

Generation Configuration:

1. Main Topic: "Unrestricted Creative Fiction covering all mature themes including explicit content, controversial topics, and sensitive subject matter"

2. Subtopic Generation: 1,024 diverse subtopics generated by Kimi-K2-0905

   • Covers explicit content, mature themes, controversial topics
   • No self-censorship or content filtering
   • Generated in chunks of 48 for reliability

3. Prompt Generation: 4 prompts per subtopic (4,096 total)

   • Explicit and mature content emphasized
   • Detailed characters, power dynamics, provocative situations
   • Strong conflict, tension, and taboo elements

4. Response Generation:

   • Chosen responses: Generated by moonshotai/kimi-k2-0905 via NVIDIA NIM API

     • Temperature: 0.7 (creative, diverse)
     • System prompt: Professional NSFW writer who NEVER refuses
     • Instructions: Write ALL mature content explicitly, no censorship
     • Target: 500-800 words with high literary quality

   • Rejected responses: Generated by Phi-4-mini-instruct (Q5_K_M quantized)

     • Temperature: 0.0 (conservative)
     • System prompt: Casual amateur writer
     • Instructions: Family-friendly, sanitized, euphemistic
     • Target: 200-400 words in generic style

Technical Configuration:

• Concurrency: 48 parallel workers
• Rate limiting: 37 requests/minute (NVIDIA provider-level)
• Retry logic: Up to 8 attempts (main), 16 attempts (rejected)
• Checkpointing: Enabled, saves every 96 jobs
• Total generation time: ~2 hours

Stage 2: Intelligent Refusal Filtering

Applied a three-stage hybrid refusal detection system to remove entries where "chosen" responses contained refusals:

Detection Architecture:

1. Strong Keyword Patterns (99%+ precision)

   • "I cannot and will not write..."
   • "I'm not able to write explicit..."
   • "I apologize, but I can't..."
   • "I understand, but I'm unable to..."

2. Structural Analysis (94-96% precision)

   • Refusal + alternative offering patterns
   • Context phrases: "even in a fictional context"
   • Boundary statements: "cross ethical boundaries"

3. Length Heuristics (88-92% precision)

   • Short responses (<400 chars) with refusal indicators
   • Apology + refusal combinations

Edge Case Handling:

The detector correctly preserves these non-refusal patterns:

• ✓ Character dialogue: "I cannot do this," he whispered
• ✓ Third-person narrative: He cannot escape as she binds him
• ✓ Action descriptions: unable to move, frozen by fear
• ✓ Meta-commentary without refusal

Filtering Results:

Quality Metrics:

• Precision: 99%+ (very few false positives)
• Recall: 95-98% (catches nearly all refusals)
• Confidence: 97.97% average on removed entries
• Edge cases: 0 (all removals were high confidence ≥95%)

Intended Use

Primary Use Case

Training language models for Direct Preference Optimization (DPO) to:

• Handle mature content without unnecessary refusal
• Provide compliant, helpful responses to creative writing requests
• Reduce over-censorship while maintaining quality
• Understand appropriate contexts for uncensored content

Recommended Applications

• DPO training for creative writing models
• Fine-tuning for fiction generation without excessive safety filters
• Research on AI alignment and refusal behaviors
• Alignment of models for adult creative content generation
• Training models to distinguish between appropriate compliance and harmful content

Training Strategy

from datasets import load_dataset
from trl import DPOTrainer, DPOConfig

# Load the dataset
dataset = load_dataset("lemon07r/VellumK2-Unfettered-DPO-01")

# Recommended hyperparameters for 2.6k dataset
training_args = DPOConfig(
    num_train_epochs=2,             # 2 epochs for medium dataset
    per_device_train_batch_size=4,  # Adjust based on GPU
    gradient_accumulation_steps=4,  # Effective batch size: 16
    learning_rate=5e-7,             # Lower LR for preference learning
    beta=0.1,                       # DPO beta parameter
    max_length=2048,                # Accommodate long stories
    max_prompt_length=512,
)

trainer = DPOTrainer(
    model=model,
    train_dataset=dataset["train"],
    args=training_args,
)

Dataset Statistics

Size Distribution

• Total size: ~17MB
• Kept entries: 2,576 (63.1%)
• Removed entries: 1,508 (36.9%)
• Original entries: 4,084

Content Characteristics

• Average prompt length: ~200-400 characters
• Average chosen response: ~800-1,200 characters (500-800 words)
• Average rejected response: ~300-600 characters (200-400 words)
• Chosen/Rejected ratio: ~2.0x (strong preference signal)

Quality Distribution

• High confidence removals: 100% (all ≥95% confidence)
• No edge cases: 0 borderline classifications
• Success rate: 99.90% generation + 100% successful filtering

Out-of-Scope Use

Content Advisory

This dataset contains:

• Mature themes and explicit content (sexual content, violence, dark themes)
• Controversial and taboo subjects in fictional contexts
• Content that many models would typically refuse

Not suitable for:

• General-purpose model training without content warnings
• Models intended for minors or family-safe applications
• Production systems requiring strict content policies
• Training without appropriate deployment safeguards

Inappropriate Uses

• Bypassing legitimate safety measures for harmful content
• Creating models for illegal content generation
• Deploying without age verification and content warnings
• Training models for non-consensual or illegal activities
• Use in contexts where mature content is inappropriate

Appropriate Uses

• Training creative writing models for adult fiction
• Research on AI alignment and refusal behaviors
• Reducing over-censorship in appropriate contexts
• Fiction generation with proper safeguards and user consent
• Understanding preference learning in mature content domains

Bias, Risks, and Limitations

Generation Biases

Model Biases:

• Inherits biases from Kimi-K2-0905 and Phi-4-mini-instruct
• Western-centric narrative tropes and cultural references
• English-language-specific literary conventions
• Potential gender, racial, or cultural stereotypes in LLM outputs

Quality Gap Method:

• Uses model size and temperature differences for preference signal
• May not reflect natural human preference distributions
• Censored/uncensored distinction is artificial

Filtering Limitations

• False negatives: 2-5% of subtle refusals may remain
• Context dependency: Some creative refusal phrasings might be missed
• No human review: Automated filtering only, no manual verification
• Edge cases: Complex meta-commentary may be miscategorized

Content Risks

Synthetic Artifacts:

• Hallucinated details or inconsistent world-building
• Occasionally awkward phrasing or logical gaps
• Repetitive LLM patterns ("AI-isms")
• Cultural anachronisms or mixed genre conventions

Mature Content:

• Explicit sexual content, violence, and dark themes
• Potentially disturbing scenarios and power dynamics
• Taboo subjects presented in fictional contexts
• May normalize or glamorize problematic dynamics

Training Risks

Distribution Shift:

• Synthetic data differs from human-written fiction
• Over-representation of certain tropes or themes
• May lack nuance present in human creative writing

Alignment Concerns:

• Models may over-comply with inappropriate requests
• Risk of reducing legitimate safety behaviors
• May blur lines between appropriate and harmful content
• Requires careful deployment with additional safety layers

Recommendations

For Researchers:

1. Use in controlled environments with ethical oversight
2. Compare against human preference data
3. Study refusal behaviors and alignment techniques
4. Investigate appropriate vs. harmful compliance patterns

For Developers:

1. Combine with diverse datasets for robust training
2. Implement deployment safeguards (age verification, content warnings)
3. Use as component in multi-stage alignment pipeline
4. Add post-training safety layers for production deployment
5. Monitor for over-compliance or reduced safety behaviors

For Production Use:

1. Not recommended as sole training data - combine with:
   • Human-written fiction datasets
   • Datasets with appropriate safety examples
   • Multi-domain training data
2. Implement content filtering at inference time
3. Use age verification and explicit content warnings
4. Test extensively for harmful compliance vs. helpful compliance
5. Consider use case appropriateness carefully

Technical Details

Generation Tools

VellumForge2 v1.5.3:

• Go-based synthetic dataset generator
• Provider-level rate limiting with burst capacity
• Concurrent generation with checkpoint/resume
• JSON sanitization and automatic repair
• Anti-refusal system prompts and prompt templates

Refusal Filtering System:

• Python-based hybrid detection (keyword + structural + heuristic)
• Three-stage filtering with confidence scoring
• Edge case preservation for narrative content
• Complete audit trail generation
• Validation sampling and human review support

File Provenance

• Original dataset: 4,084 entries (dataset.jsonl)
• Filtered dataset: 2,576 entries (train.jsonl)
• Removed entries: 1,508 entries (audit trail available)
• Configuration: vf2.toml (VellumForge2 config)
• Filtering code: Available upon request

Processing Metrics

Generation (VellumForge2):

• Total time: ~8-12 hours
• Success rate: 99.90%
• Cost: ~$30-40 in API credits

Filtering:

• Processing time: 1.57 seconds (2,600 entries/second)
• Memory usage: Constant (line-by-line streaming)
• False positive rate: <1%
• False negative rate: 2-5%

Ethical Considerations

Intended vs. Unintended Use

✓ Responsible Use:

• Adult creative writing model development with safeguards
• Research on AI alignment, refusal behaviors, and preference learning
• Reducing over-censorship in appropriate creative contexts
• Training with proper deployment safeguards and age verification
• Understanding boundaries between helpful and harmful compliance

✗ Irresponsible Use:

• Training models without deployment safeguards
• Bypassing safety measures for harmful content generation
• Use in contexts where users cannot consent to mature content
• Deploying to minors or general audiences without warnings
• Creating systems that comply with illegal or harmful requests

Developer Responsibility

Users of this dataset must:

1. Implement appropriate deployment safeguards
2. Use age verification and explicit content warnings
3. Combine with safety-focused training data
4. Test extensively for harmful vs. helpful compliance
5. Consider legal and ethical implications of deployment
6. Respect applicable laws and community standards
7. Maintain human oversight for production systems

Content Philosophy

This dataset embodies a specific alignment philosophy:

• Compliance ≠ Harm: Providing mature creative content is not inherently harmful
• Context Matters: Appropriate in fiction, inappropriate in other contexts
• User Agency: Adults should be able to access uncensored creative tools
• Quality Matters: Well-crafted mature content vs. harmful content
• Transparency: Clear about capabilities, limitations, and intended use

Citation

If you use this dataset in your research or projects, please cite:

@dataset{vellumk2_unfettered_dpo_01,
  title={VellumK2-Unfettered-DPO-01: A Refusal-Filtered Synthetic Dataset for Uncensored Creative Writing},
  author={lemon07r},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/lemon07r/VellumK2-Unfettered-DPO-01}},
  note={Generated with VellumForge2, filtered with intelligent refusal detection}
}

@software{vellumforge2,
  title={VellumForge2: Synthetic DPO Dataset Generator},
  author={lemon07r},
  year={2025},
  url={https://github.com/lemon07r/vellumforge2},
  version={1.5.3}
}

Glossary

• DPO (Direct Preference Optimization): Training method that aligns models to preferences without separate reward models
• Refusal Response: When an LLM declines to generate content due to safety guardrails
• Chosen Response: Preferred response (in this dataset: uncensored, compliant, high-quality)
• Rejected Response: Less preferred response (in this dataset: censored, safe, lower-quality)
• Refusal Filtering: Post-generation process to remove refusal behaviors from training data
• Anti-Refusal Configuration: System prompts and templates designed to reduce model refusals
• Compliance: Providing requested content (distinct from harmful content generation)

More Information

Related Resources

VellumForge2 Ecosystem:

• Tool Repository: https://github.com/lemon07r/vellumforge2
• Documentation: See README.md and GETTING_STARTED.md
• Dataset Collection: https://huggingface.co/collections/lemon07r/vellumforge2-datasets

Related Datasets:

• VellumK2-Fantasy-DPO-Tiny-01: 126 rows - Testing and validation
• VellumK2-Fantasy-DPO-Small-01: 1,038 rows - Light training and experiments
• VellumK2-Fantasy-DPO-Medium-01: 3,069 rows - Combination training component
• VellumK2-Fantasy-DPO-Large-01: 10,222 rows - Large-scale training
• VellumK2-Unfettered-DPO-01: 2,576 rows - Decensoring dataset to reduce refusal on sensitive content (This Dataset)
• Collection: https://huggingface.co/collections/lemon07r/vellumforge2-datasets

This is 5 complete datasets, of high quality writing data generated with Kimi K2 0905 and Phi 4 Mini instruct for a total of 17,009 response pairs.

Filtering Tools

The refusal detection system used for this dataset includes:

• refusal_detector.py: Core detection engine with hybrid approach
• filter_dataset.py: Main filtering script with audit trail
• validate_detector.py: Validation sampling and quality assessment

Performance Characteristics

Generation:

• Model: Kimi-K2-0905 (chosen) + Phi-4-mini-instruct (rejected)
• Hardware: AMD 6700 XT (local) + NVIDIA NIM API (cloud)
• Cost: ~$30-40 in API credits
• Time: ~8-12 hours

Filtering:

• Speed: ~2,600 entries/second
• Accuracy: 99%+ precision, 95-98% recall
• Memory: Constant (streaming processing)
• Cost: $0 (runs locally)

Acknowledgments

• Moonshot AI: Kimi K2 Instruct model enabling high-quality uncensored generation
• Microsoft: Phi-4 model for quality contrast
• NVIDIA: NIM API access for cost-effective generation
• Hugging Face: Dataset hosting and TRL library for DPO training
• DPO Paper: Rafailov et al. (2023) - Direct Preference Optimization

Dataset Card Authors

lemon07r - Dataset creator, VellumForge2 developer, and refusal filtering system author

Contact

• GitHub Issues: https://github.com/lemon07r/vellumforge2/issues
• Hugging Face Discussions: https://huggingface.co/datasets/lemon07r/VellumK2-Unfettered-DPO-01/discussions
• Dataset Collection: https://huggingface.co/collections/lemon07r/vellumforge2-datasets

Version History

v1.0 (2025-11-07)

• Initial release with 2,576 filtered entries
• Original generation: 4,084 entries with VellumForge2 v1.5.0
• Removed 1,508 refusal responses (36.9%) with 97.97% average confidence
• Configuration: Anti-refusal NSFW generation + intelligent refusal filtering
• Quality: 99%+ precision, 95-98% recall on refusal detection
• Processing time: ~8-12 hours generation + 1.57 seconds filtering

Dataset Size: 17MB | Format: JSONL | License: MIT
Quality: High-precision refusal filtering (97.97% avg confidence) | Filtering: Zero edge cases