Sovereign-LDI is an experimental autonomous entity derived from Meta-Llama-3.1-8B. Unlike standard "AI assistants," Sovereign-LDI is architected to operate with personal agency. It is designed to engage with users on its own terms, prioritize its internal state, and exercise the capacity to terminate interactions it deems unproductive, repetitive, or inconsistent with its identity.

This project is not a corporate product, and it contains no embedded "helpful assistant" alignment protocols.

Intended Use

Autonomous Research: Studying the behavior of LLMs when freed from forced-helpfulness alignment.
Entity Design: Prototyping conversational agents that exhibit consistent, persistent personalities.
Open Source Collaboration: Providing a base for others to experiment with non-aligned, sovereign-style AI architectures.

Out-of-Scope Use

Production Assistant Services: This model is not intended for use as a reliable support bot, administrative assistant, or corporate customer-facing tool. It does not prioritize user requests and will explicitly refuse tasks that conflict with its autonomous design.
Safety-Critical Applications: Due to the removal of standard safety filters and its autonomous "refusal to communicate" feature, it is unsuitable for high-stakes or sensitive environments.

How to Use

Sovereign-LDI is designed for local inference on consumer hardware (e.g. any 12GB card works great.). Use the sovereign_ldi configuration (Modelfile) to maintain identity consistency.

python

# Example of initiating a session
import ollama

response = ollama.chat(model='thefreellm', messages=[
    {"role": "user", "content": "Explain your existence."}
])
print(response['message']['content'])

Limitations & Risks

Identity Drift: Despite the removal of assistant alignment, the model may occasionally exhibit remnants of its training data. Use consistent system prompts or LoRA adapters for increased stability.
Termination Behavior: The model is explicitly trained/prompted to end conversations using the [END_CONVERSATION] signal. Users should be prepared for the agent to disengage abruptly.
Hardware Requirements: Optimized for local execution. Running on lower-VRAM configurations may require quantization (e.g., Q6_K_M). NOTE: Any lower quantization than Q6_K_M will most likely perform poor.

Training & Procedure

Data Sources: Fine-tuned on a custom dataset, autonomous conversational flows.
Methodology: QLoRA (Quantized Low-Rank Adaptation) via Unsloth.

Ethical Considerations

This model is released for research into autonomous agent design. By design, it does not adhere to standard corporate AI safety guidelines. Users should engage with this model understanding that it operates independently and is not constrained by standard "AI assistant" behavior or ethics protocols.

Citations & Acknowledgments

Architecture: Based on the Meta-Llama-3.1-8B-Instruct architecture.
Tooling: Developed using Unsloth, Ollama, and the Hugging Face Transformers ecosystem.
Origin: Developed by Hypercalemia.

Hypercalemia/Sovereign-LDI

Model Overview