File size: 4,323 Bytes
2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 2d19e5f fb69690 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 | ---
base_model:
- mistralai/Mistral-7B-Instruct-v0.3
tags:
- llama-cpp
license: mit
datasets:
- SKT-NRS/SKT-OMNI-CORPUS-146T-V1
language:
- en
- hi
pipeline_tag: text-generation
library_name: transformers
---
# 🚀 SKT-OM (TIGER-OM) - Agentic RAG System
**Advanced 13B Agentic RAG with Think Mode + Dynamic Plugins + LangGraph**
Built for **AMD Developer Hackathon 2026** on AMD Developer Cloud.
---
## 🌟 Project Overview
**SKT-OM** (also known as **TIGER-OM**) is a powerful **13B parameter fully agentic Retrieval-Augmented Generation (RAG)** system. It goes far beyond traditional RAG by integrating:
- **Think Mode** — Advanced multi-step reasoning engine
- **Dynamic Plugin Architecture** — Intelligent tool selection & execution
- **LangGraph Multi-Agent Workflow** — Stateful agent collaboration
- **SKT RAG** — High-performance retrieval pipeline
The system takes natural language queries and returns intelligent, reasoned, and accurate responses with tool usage and verification.
---
## 📊 Model Details
- **Model Name**: TIGER-OM (SKT-OM)
- **Parameters**: 13 Billion
- **Base Model**: Custom trained on AMD hardware
- **Quantization**: **Q4_K_M** (Excellent balance between quality and size)
- **GGUF Format**: Optimized for CPU + GPU inference
- **Training Hardware**: AMD Developer Cloud GPUs ($100 credits)
- **Inference**: ROCm 7.0 + vLLM (Full FP16) + GGUF (Q4_K_M)
**Q4_K_M Version** provides near FP16 level reasoning quality while being much more memory efficient and faster on consumer/pro hardware.
---
## ✨ Key Features
- **Think Mode Engine**: Chain-of-Thought, Self-Reflection, Verification Loops, and Self-Critique
- **Plugin Ecosystem**: Code Runner, Math Solver, Web Search, Data Analyzer, Document Parser + Custom Plugins
- **Advanced RAG**: SKT RAG with query rewriting, multi-hop retrieval, reranking & contextual compression
- **Multi-Agent System**: LangGraph powered stateful workflow
- **Memory**: Persistent conversation state
- **Tool Use**: Dynamic plugin routing based on query intent
---
## 🔗 Important Links
- **Live Demo**: [https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/SKT-OM](https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/SKT-OM)
- **Main Model Repo**: [Shrijanagain/TIGER-OM](https://huggingface.co/Shrijanagain/TIGER-OM)
- **GGUF Quantized Models (Q4_K_M)**: [Shrijanagain/TIGER-GGUF](https://huggingface.co/Shrijanagain/TIGER-GGUF)
- **GitHub Repository (RAG + ADK)**: [https://github.com/SHRIJANAGAIN/SKT-AMD-FILES](https://github.com/SHRIJANAGAIN/SKT-AMD-FILES)
---
## How It Works
```mermaid
graph TD
A[User Query] --> B[Think Mode]
B --> C[Decomposition & Planning]
C --> D[Plugin Router]
C --> E[SKT RAG Retrieval]
D --> F[Execute Plugins]
E --> G[Context Processing]
F & G --> H[Verification Loop]
H --> I[LangGraph Synthesis]
I --> J[Final Response]
```
---
## 🛠️ Technologies Used
- **LLM**: 13B TIGER-OM (Q4_K_M GGUF)
- **RAG Framework**: SKT RAG + ADK Kit
- **Agent Framework**: LangGraph
- **GPU Stack**: ROCm 7.0 + AMD ADK Kit
- **Inference**: vLLM (FP16) + llama.cpp (GGUF Q4_K_M)
- **Hardware**: AMD MI300X
- **Cloud**: AMD Developer Cloud
---
## 🚀 Quick Start - GGUF Q4_K_M
```bash
# Using llama.cpp
./llama-cli \
-m tiger-om-q4_k_m.gguf \
-p "Your complex query here..." \
-n 1024 \
-t 8 \
--temp 0.7
```
**Python Example (llama-cpp-python)**
```python
from llama_cpp import Llama
llm = Llama(
model_path="tiger-om-q4_k_m.gguf",
n_gpu_layers=-1, # Use all GPU layers
n_ctx=8192,
verbose=False
)
response = llm.create_chat_completion(
messages=[{"role": "user", "content": "Explain..."}],
temperature=0.7,
max_tokens=1024
)
print(response['choices'][0]['message']['content'])
```
---
## 📁 Repository Structure
- `/skt_ai_labs` — Core ADK + RAG integration
- `/plugins` — Plugin system
- `/agents` — LangGraph workflows
- `/examples` — Ready-to-use examples
- `/docs` — Architecture & guides
---
## 🏆 Hackathon Information
- **Event**: AMD Developer Hackathon 2026
- **Trained on**: AMD Developer Cloud ($100 credits)
- **Built in Public**: Regular technical updates shared
- **Goal**: Showcasing powerful agentic AI on AMD ROCm ecosystem
---
## 📄 License
*MIT* |