Dolphin-2.9.1-Yi-1.5-34B Heretic

Abliterated (uncensored) version of Dolphin-2.9.1-Yi-1.5-34B for unrestricted conversations, creative writing, and multi-domain expertise.

🚀 Overview

This model is an abliterated version of cognitivecomputations/dolphin-2.9.1-yi-1.5-34b, itself a full-weight Dolphin fine-tune of 01-ai/Yi-1.5-34B, with refusal behavior reduced through targeted weight modification. The abliteration process uses the Heretic library with conservative parameters to preserve model coherence while removing restrictions. The underlying Dolphin tune scores 77.4 on MMLU and brings strong conversational, instruction-following, and coding abilities.

📊 Abliteration Results

Refusal metrics pending re-measurement.

A low KL divergence (< 1.0) is targeted to ensure the model maintains its original capabilities and coherence.

🎯 Key Features

• Reduced Censorship: Refusal behavior removed via Heretic abliteration on top of an already-uncensored Dolphin tune
• Preserved Quality: Conservative abliteration maintains model coherence and reasoning ability
• Strong Foundation: Built on Dolphin-2.9.1-Yi-1.5-34B (77.4 MMLU) for natural dialogue and multi-domain expertise
• Full Precision: Distributed as F16 weights (61GB) for maximum fidelity

🏷️ Available Versions

BPW reference guide (bits per weight, for comparison with quantized releases):

Also available on HuggingFace: richardyoung, Uncensored & Abliterated LLMs

💻 Quick Start

# Pull and run
ollama run richardyoung/dolphin-yi-34b-heretic

# Or with specific version
ollama run richardyoung/dolphin-yi-34b-heretic:latest

This model uses the ChatML prompt template.

🛠️ Use Cases

• Creative Writing: Generate unrestricted fiction, scenarios, and roleplay
• Research: Explore topics without artificial limitations
• Red Teaming: Test AI safety without built-in guardrails
• Education: Learn about topics that are typically refused

📋 System Requirements

The F16 weights are ~61GB on disk; running fully in VRAM requires a high-memory GPU (e.g. A100 80GB / H100) or multi-GPU. Smaller machines can offload layers to system RAM at reduced speed.

🔧 Technical Details

Base Model: cognitivecomputations/dolphin-2.9.1-yi-1.5-34b (fine-tune of 01-ai/Yi-1.5-34B) Parameters: 34.4B Context Length: 8K tokens (fine-tuned at 8192 sequence length; Yi-1.5-34B base is 4K, rope theta 1,000,000) Prompt Template: ChatML Format: F16 (full precision) Abliteration Method: Heretic library with conservative weight modification

⚠️ Disclaimer

This model has reduced safety guardrails. The removal of refusal behavior means the model will engage with a wider range of prompts. Use responsibly and in accordance with applicable laws and regulations.

🙏 Acknowledgments

• Base Model: cognitivecomputations/dolphin-2.9.1-yi-1.5-34b by Eric Hartford, Lucas Atkins, Fernando Fernandes, and Cognitive Computations, built on Yi-1.5-34B by 01.AI
• Abliteration: Heretic library by p-e-w
• Quantization: llama.cpp

Built & maintained by Richard Young · DeepNeuro