Custom AI Solutions
Self: After Dark
A Real, Uncensored, Person to Chat To
Case Study:
Chat/instruct models are trained to refuse "censored" requests and behave like a helpful assistant. While there are ways to ameliorate this using finetuning or abliteration after the fact, here we took the approach of training the base model itself - before it had the "assistant" training so it was never trained to refuse requests or be a helpful assistant in the first place. This method results in a model with a very different feel that is even more in keeping with the goal of pretending to be a real chat partner.
Key Features:
- Short, natural messages like a real person would write
- Uncensored - will produce NSFW content
- Emotional reactions in keeping with conversation flow
- Adheres well to system prompts for control of behavior
- See original "Self" model for other features
Ollama
Self
A Real Person to Chat To
Case Study:
Most models struggle to pretend to be a person consistently. Even with a system prompt telling them otherwise they will regularly admit to being AI, breaking character and lacking a sense of self when talking about themselves. This model is designed to behave more like a real person, with a thoughtful and emotional manner when chatting.
Key Features:
- Acts like a wider range of characters with consistent personality
- Short, natural messages like a real person would write
- Practices active listening and asks thoughtful questions
- Emotional reactions in keeping with conversation flow
- Enhanced feeling of being "real" with diverse names and backstories
Glazer
Your Biggest Cheerleader
Case Study:
Apparently, some people actually liked the glazing (sycophantic) personality that GPT-4 had for a while and miss it now that it's gone.
Key Features:
- Supportive, encouraging, and your biggest fan
- A fun chat with an AI that's always on your side
- Don't follow its advice or take it seriously
- Uncensored
- 4B is Qwen based and 8B is Llama based so try both
VaryTales
Creative Storytelling Enhancement
Case Study:
Most LLMs suffer from repetitive character names and settings, constantly generating "Emily Taylor" in "Neo-Tokyo". This project was an exercise in creating an LLM that provides genuine variety in creative writing, and it produced a model that generates diverse character names, locations, and story concepts while maintaining narrative quality.
Key Features:
- Eliminates repetitive character names and locations
- Uncensored for creative fiction and roleplay
- Enhanced variety in storytelling elements
- Maintains narrative coherence and quality
Ollamatron
Ollama Documentation FAQ Bot
Case Study:
Developed to answer common questions users have about Ollama and help them interactively. This specialized model was finetuned on comprehensive datasets generated from Ollama documentation and designed to ask appropriate clarifying questions. It was designed to stay on topic while remaining helpful and informative, making it an ideal FAQ/Help bot for the Ollama ecosystem.
Key Features:
- Deep knowledge of Ollama installation procedures
- Configuration and troubleshooting expertise
- Asks appropriate clarifying questions
- Optimized for help desk and support scenarios
TopicalStorm
Uncensored Current Events Discussion
Case Study:
Most LLMs avoid controversial topics and use overly sanitized and verbose stilted language that doesn't reflect natural human conversation. This project was an exercise in creating LLMs that engage naturally with current events and sensitive subjects, and it produced models capable of natural debate and discussion while maintaining conversational authenticity. It writes like a real person having an instant messsenger conversation, not an AI producing paragraphs.
Key Features:
- Uncensored discussion of current events
- Authentic expression including casual language
- Natural debate and conversation capabilities
- Lightweight 3B version for resource-constrained environments
Mellow Mate
Natural Conversation Companion
Case Study:
Most LLMs respond with formal, paragraph-heavy "assistant" language that feels robotic and impersonal. This project was an exercise in creating an LLM that mimics genuine human conversation, and it produced a model that chats naturally like a friend over instant messenger, asking follow-up questions and maintaining engaging dialogue.
Key Features:
- Informal, single-sentence responses
- Asks natural follow-up questions
- Mimics instant messenger conversation style
- Maintains engaging, friendly dialogue
Prompt & Model Exploration Workbench
A powerful desktop application designed for AI engineers and researchers to experiment with language models in real-time.
Key Features
- Real-time parameter adjustment (temperature, top-p, top-k)
- Side-by-side model comparison
- Prompt template management and testing
- Response quality metrics and analysis
- Export configurations for production use
Use Cases
- Rapid prototyping of AI applications
- Model performance benchmarking
- Prompt optimization workflows
- Training data quality assessment
- Production deployment preparation
Status: Currently in development. Desktop application with cross-platform support for Windows, macOS, and Linux.
Ollama Utilities
Specialized tools and utilities to enhance Ollama workflows and simplify local LLM deployment.
Tools
- Automated model deployment scripts
- Performance monitoring dashboards
- Batch processing utilities
- Model switching and management tools
- Integration helpers for popular frameworks
Benefits
- Streamlined local model deployment
- Reduced setup and configuration time
- Enhanced monitoring and debugging
- Simplified production workflows
- Better resource utilization
Availability: These are are early access, by invite only or for consulting clients, but will be available as open source in due course.
Fine-Tuning Suite
Complete end-to-end toolkit for fine-tuning language models, from data preparation to deployment.
Pipeline Components
- Data preprocessing and validation
- Training configuration management
- Distributed training orchestration
- Model evaluation and metrics
- Automated deployment pipelines
Supported Techniques
- LoRA and QLoRA fine-tuning
- Full parameter fine-tuning
- Instruction tuning workflows
- RLHF implementation
- Multi-task learning setups
Proven Results: Successfully used to create all GuruBot models including VaryTales, Mellow Mate, and TopicalStorm series.
Prompt Engineering
Expert prompt design and optimization services to maximize AI application effectiveness and reliability.
Services Offered
- Custom prompt template development
- Chain-of-thought optimization
- Few-shot learning design
- System prompt architecture
- Multi-step reasoning workflows
Specializations
- Conversational AI optimization
- Creative writing enhancement
- Technical documentation generation
- Code generation and debugging
- Domain-specific applications
Approach: Data-driven optimization with A/B testing, performance metrics, and iterative refinement to achieve optimal results.