Dolphin 3.0 Llama3.2 3B

Overview

Dolphin 3.0 Llama 3.2 3B is an instruct-tuned generative language model developed by Cognitive Computations, with key contributions from Eric Hartford, Ben Gitter, and BlouseJury. As part of the Dolphin 3.0 Collection, this model emphasizes user autonomy and adaptability, offering capabilities in natural language understanding, programming, mathematics, and agentic functions. Notably, it features an uncensored design, allowing full user control over system prompts, alignment, and tone. Dolphin 3.0 Llama 3.2 3B builds on the Llama 3.2 3B base model, with 3.2 billion parameters and a focus on composable alignment for diverse user needs.

Technical Architecture and Capabilities

Dolphin 3.0 Llama 3.2 3B is constructed atop the Llama 3.2-3B architecture, leveraging 3.21 billion parameters and utilizing the BF16 tensor type for efficient computation. The model is trained to function as a general-purpose language model, with proficiency in coding, mathematical reasoning, function calling, agentic operations, and broad conversational tasks. Its instruct-tuned setup ensures adaptability to a wide spectrum of user-defined instructions and operational domains.

A defining feature is its uncensored approach, giving users unrestricted control over system prompts and the model's behavioral alignment. Unlike many mainstream AI systems, which tightly govern prompt structure and alignment—sometimes leading to inconsistent behavior or unannounced updates—Dolphin 3.0 Llama 3.2 3B is deliberately designed to delegate ethical and operational responsibility to the user. Developers maintain that this approach grants maximum flexibility for model alignment, particularly useful for specialized, creative, or privacy-focused applications. Further discussion of the underlying rationale can be found in Eric Hartford’s examination of uncensored models.

Training Procedure and Data

In developing Dolphin 3.0 Llama 3.2 3B, the team employed a combination of open-source datasets and specialized training methodologies. The training corpus includes resources from OpenCoder-LLM, Microsoft, NousResearch, AI-MO, Allen Institute for AI, HuggingFaceTB, and M-a-p. Augmentation with Deepseek-V3 accelerated dataset expansion, while RLHFlow provided reward models to refine data quality through advanced filtering.

A core aspect of training is the removal of "refusals" and explicit biases from instruction datasets, a process central to the production of uncensored models. Instruction datasets often carry imprints of refusal behaviors, particularly learned from models highly aligned to specific ethical or policy frameworks (such as ChatGPT, Claude, or Gemini). The Dolphin 3.0 approach systematically filters out such content to ensure objective and comprehensive outputs in response to diverse prompts. This methodology is detailed in the Uncensored Models article.

Alignment Philosophy and Ethical Considerations

Dolphin 3.0 Llama 3.2 3B is structured as an uncensored model. Rather than enforcing a singular ethical alignment, it leaves prompt structure and behavioral scope entirely in the hands of the user. This contrasts with models such as ChatGPT, Claude, and Gemini, where alignment and refusal mechanisms may constrain valid use cases, introduce unpredictable changes, or affect user privacy and control. By granting composable alignment, Dolphin enables domain-specific, culturally tailored, or privacy-sensitive deployments.

This explicit design ensures that the system's owner retains responsibility for the outputs, supporting use cases across creative writing, roleplay, agentic workflows, or specialized domains that require full transparency or oversight over generated content. More details on the philosophical underpinnings are available in the Uncensored Models blog post.

Usage Patterns and Applications

Dolphin 3.0 Llama 3.2 3B is optimized for a broad array of tasks. Its instruct-tuned architecture enables sophisticated dialogue, general technical assistance, programming support, and mathematical problem-solving. The model is often used in scenarios that demand custom alignment, such as implementing specific character roles, enforcing domain-specific rules, or preserving user privacy in sensitive domains.

The model operates using the ChatML formatting protocol, allowing users to define system prompts with granular specificity. For example, a user may configure the model to act exclusively as a coding assistant for a particular language, or as a conversational agent with a stance tailored to specific worldviews or professional norms. This flexibility is exemplified in the output samples, which demonstrate the model’s capacity for role adaptation and detailed technical generation.

Limitations and Considerations

As an uncensored language model, Dolphin 3.0 Llama 3.2 3B does not incorporate behavioral refusal, alignment constraints, or safety filters by default. While this design increases flexibility, it also places the onus for downstream application and ethical stewardship entirely on the user. The model is capable of producing responses on a wide range of topics—including those that would typically trigger refusal in more strictly aligned architectures. Users must ensure that deployments comply with relevant legal and ethical guidelines.

Model evaluation metrics for Dolphin 3.0 Llama 3.2 3B are marked as "to be determined," and continuous updates to the Dolphin 3.0 Collection may introduce incremental improvements and variants.

Development Background and Model Family

Dolphin 3.0 Llama 3.2 3B belongs to a lineage of uncensored models established by Eric Hartford and collaborators. Predecessors in the series include WizardLM-30B-Uncensored, WizardLM-13B-Uncensored, WizardLM-7B-Uncensored, and Wizard-Vicuna-13B-Uncensored. Regular updates and refinements are summarized in the Dolphin 3.0 Collection, with notable milestones such as the update on February 7. The development of the original Dolphin models was supported in part by a grant from Andreessen Horowitz, dedicated to advancing open-source AI research.

Licensing and Open-Source Principles

While the model’s licensing details are not explicitly stated, the overall development philosophy is rooted in open-source ideals, prioritizing transparency, user control, and the right to define operational alignment. This ethos aligns with broader trends in open generative AI, reflecting growing recognition of the importance of customizable and user-governed language technologies.

Helpful External Resources

• Dolphin 3.0 Collection on Hugging Face
• Cognitive Computations official Discord
• Uncensored Models article by Eric Hartford
• Eric Hartford’s Hugging Face profile
• Ben Gitter’s Hugging Face profile
• BlouseJury’s Hugging Face profile
• Cognitive Computations Hugging Face profile
• OpenCoder-LLM on Hugging Face
• NousResearch on Hugging Face
• AI-MO on Hugging Face
• AllenAI on Hugging Face
• HuggingFaceTB on Hugging Face
• M-a-p on Hugging Face
• RLHFlow on Hugging Face
• Safetensors documentation
• Andreessen Horowitz grant for open-source AI
• Political bias in ChatGPT (Brookings Institute)
• ChatML format details