YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
IRIS 1.1 TEST 🧠
«Every version is a step toward a more capable IRIS.»
IRIS is a personal AI model and assistant developed by Revanth Annapureddy under RRAD — Revanth Research and Development.
IRIS 1.1 TEST is an experimental fine-tuned language model built on top of Google Gemma 3 270M. The project focuses on creating a personalized AI with its own identity, personality, conversational behavior, and future agentic capabilities.
👨💻 Developer
Revanth Annapureddy
Organization: RRAD — Revanth Research and Development
IRIS is developed as part of Revanth's personal research and development work in artificial intelligence, local LLMs, fine-tuning, AI agents, and application development.
🤖 About IRIS
IRIS 1.1 TEST is a personalized AI model designed to:
- Have its own identity and personality
- Provide general conversation and assistance
- Interact in a friendly and entertaining way
- Understand information about its creator and project
- Provide jokes, riddles, puzzles, and thrill-based conversations
- Serve as the foundation for future AI applications
- Experiment with local and offline AI technology
IRIS is not trained from scratch. Its language-model foundation is Gemma 3 270M, while its identity, datasets, fine-tuning, and application design are developed as part of the IRIS project.
🧬 Model Information
Property| Details Model| IRIS 1.1 TEST Base Model| Gemma 3 270M Developer| Revanth Annapureddy Organization| RRAD Training Method| Fine-tuning Fine-tuning Approach| LoRA / QLoRA Output Format| GGUF Quantization| Q8_0 Model Type| Personalized LLM Status| Experimental / Testing
🔬 Fine-Tuning
IRIS 1.1 TEST was created by fine-tuning Gemma 3 270M with custom datasets designed specifically for the IRIS personality and behavior.
Fine-Tuning Pipeline
Gemma 3 270M │ ▼ Custom IRIS Datasets │ ▼ LoRA / QLoRA Fine-Tuning │ ▼ Trained Model / Adapter │ ▼ Model Conversion │ ▼ GGUF │ ▼ Q8_0 Quantization │ ▼ IRIS 1.1 TEST
Training Objective
The main objective was not to create a completely new foundation model, but to customize the behavior of Gemma into a personalized assistant with:
- IRIS identity
- Creator awareness
- Project knowledge
- Friendly personality
- Humorous responses
- General conversational ability
- Riddle and puzzle interaction
- Future AI-agent compatibility
📚 Datasets
IRIS 1.1 TEST uses multiple datasets with different purposes.
- General Conversation Dataset
Used to improve:
- General knowledge
- Question answering
- Natural conversation
- Instruction following
- Helpful responses
- Entertainment / Thrill Dataset
Designed to make IRIS more engaging.
It contains examples involving:
- Jokes
- Puns
- Riddles
- Brain teasers
- Paradox questions
- Wordplay
- Guessing games
- Interactive challenges
The goal is for IRIS to do more than simply answer questions and instead create an enjoyable conversation.
- IRIS Identity Dataset
A custom identity dataset created specifically for IRIS.
It teaches information such as:
- Who IRIS is
- Who developed IRIS
- What RRAD is
- Why IRIS was created
- IRIS's current version
- IRIS's personality
- IRIS's mobile vision
- IRIS's desktop vision
- IRIS's development philosophy
- Future planned versions
The identity dataset contains 500 examples.
🧾 Dataset Format
The training data follows an instruction/input/output structure.
Example:
{ "instruction": "Who are you?", "input": "", "output": "I am IRIS 1.1 TEST, a personal AI model developed by Revanth Annapureddy under RRAD — Revanth Research and Development." }
Another example:
{ "instruction": "Who created you?", "input": "", "output": "I was developed by Revanth Annapureddy under RRAD — Revanth Research and Development." }
🎭 IRIS Personality
IRIS is designed to be:
- 🧠 Smart
- 😊 Friendly
- 😂 Funny
- 💬 Conversational
- ⚡ Helpful
- 🎯 Personalized
IRIS may address its creator as:
«Master or Sir»
depending on the conversation.
📱 Mobile Vision
The planned mobile version of IRIS focuses on:
- General chat
- Personal AI interaction
- Entertainment
- Jokes
- Riddles
- Thrill conversations
- Lightweight AI assistance
The goal is to eventually integrate the IRIS model into a personal Android AI application.
🖥️ Desktop Vision
The broader desktop version is planned to become a more capable AI system with features such as:
- 💻 Coding Agent
- 👁️ Vision
- 🧠 RAG
- 🛠️ Agentic Tools
- 📁 File Management
- 🌐 Browser Tools
- 🎙️ Voice
- 🧠 Memory
- 🔧 Developer Tools
These capabilities are part of the application vision and should not be confused with capabilities currently built into the IRIS 1.1 TEST language model itself.
📊 Training Result
The reported average training loss for the current experiment was approximately:
Average Loss: 1.3
The resulting model was converted to:
GGUF
with:
Q8_0
quantization.
«Note: Training loss alone does not determine model quality. Generation tests, validation loss, instruction following, hallucination rate, and runtime behavior are also important for evaluating IRIS.»
⚠️ Current Status
IRIS 1.1 TEST is an experimental version.
The current stage focuses on evaluating:
- Identity consistency
- Response quality
- Hallucination
- Random-token generation
- Instruction following
- Personality
- General conversation
- GGUF conversion quality
- Runtime compatibility
Some problems may still occur, including incorrect or unexpected responses.
These problems are considered part of the experimental development process.
🧪 Testing
Example prompts for testing IRIS:
Who are you?
Who created you?
What is RRAD?
What model is your foundation?
Why were you created?
What is your current version?
Tell me a joke.
Give me a riddle.
Ask me a paradox question.
What is your mobile vision?
What is your desktop vision?
What should happen if IRIS fails?
What are the future planned versions of IRIS?
🛠️ Technology Stack
IRIS research currently involves technologies such as:
- Python
- PyTorch
- Hugging Face
- Gemma
- LoRA / QLoRA
- GGUF
- llama.cpp
- Ollama
- Local LLM inference
- Android development
- AI Agents
- RAG
- Vector databases
🗺️ Development Roadmap
✅ IRIS 1.1 TEST
- Define IRIS identity
- Create custom identity dataset
- Fine-tune Gemma 3 270M
- Generate GGUF model
- Q8_0 quantization
- Begin model testing
- Improve hallucination
- Improve response quality
- Improve conversational consistency
🔮 IRIS 1.5 Flash-Code
Planned future version focused on stronger coding and development capabilities.
🔮 IRIS 1.5 Sol
Planned future version focused on a more capable general-purpose IRIS system.
«Future versions are planned and are not currently released.»
🧠 Development Philosophy
IRIS follows a simple philosophy:
«If IRIS fails, understand why it failed. Learn from the failure. Improve the next version.»
Every experiment provides information that can be used to make the next version better.
🎯 Long-Term Goal
The long-term goal of the IRIS project is to develop a personalized AI ecosystem that can operate locally and eventually provide:
┌─────────────────┐
│ IRIS │
│ Personal AI │
└────────┬────────┘
│
┌────────────────┼────────────────┐
│ │ │
▼ ▼ ▼
Android Desktop Local AI
│ │ │
▼ ▼ ▼
Chat AI Agent Offline LLM
│
┌──────────┼──────────┐
▼ ▼ ▼
RAG Vision Tools
👨🔬 Project Philosophy
IRIS is more than a fine-tuned model.
It is an ongoing experiment in:
- Local LLM development
- Model fine-tuning
- AI personalization
- AI application development
- Agentic AI
- Human-AI interaction
- Learning through experimentation
📜 Disclaimer
IRIS 1.1 TEST is an experimental research project.
The model is based on Gemma 3 270M and should not be represented as a model trained from scratch.
IRIS's identity, datasets, fine-tuning work, application concepts, and project direction are developed under RRAD — Revanth Research and Development.
👨💻 Creator
Revanth Annapureddy
RRAD — Revanth Research and Development
«IRIS 1.1 TEST Every version is a step toward a more capable IRIS.»
- Downloads last month
- 209
8-bit