base_model:
- microsoft/Phi-3.5-mini-instruct
Phi3.5-Ludacris.Instruct.Uncensored.GGUF is a compact uncensored instruction-tuned language model based on Microsoft’s Phi-3.5-mini-instruct architecture.
This release focuses on:
- 🔓 Reduced refusal behavior
- 🧠 Strong small-model reasoning
- ⚡ Efficient local inference
- 💻 Instruction following + coding capability
- 🧩 GGUF deployment simplicity
The model is distributed exclusively in GGUF format for fast local execution through:
- llama.cpp
- LM Studio
- KoboldCpp
- Ollama (manual import)
- text-generation-webui
- llama-cpp-python
🧬 Base Model
| Attribute | Value |
|---|---|
| Base Model | Phi-3.5-mini-instruct |
| Creator | Microsoft |
| Architecture | Transformer-based causal LLM |
| Parameter Size | ~3.8B |
| Context Length | 128K |
| Format | GGUF |
| Quantization Available | Q4_K_M only |
Microsoft designed Phi-3.5-mini-instruct as a lightweight reasoning-focused model with strong instruction-following behavior and long-context support. ([Reddit][1])
🔓 Uncensored Variant
This version was modified by Within Us AI to reduce alignment restrictions and refusal-heavy behavior found in the original Phi-3.5 release.
Community discussion around Phi-3.5 often described the original model as extremely restrictive compared to many open-weight alternatives. ([Reddit][2])
The goal of this release is to preserve:
- reasoning ability
- instruction quality
- coding usefulness
- conversational coherence
…while reducing excessive refusals and over-filtering.
⚙️ Quantization
Available Quant
| Quant | Size Class | Recommended Use |
|---|---|---|
| Q4_K_M | Balanced 4-bit quant | Best balance of quality + speed |
This repository currently includes only the Q4_K_M GGUF variant.
Q4_K_M is commonly favored in the GGUF ecosystem because it preserves strong output quality while remaining lightweight enough for consumer hardware. ([Reddit][3])
🚀 Intended Use
Ideal For
- Local AI assistants
- Offline inference
- Creative writing
- Coding assistance
- Long-context experiments
- AI research
- Unfiltered conversational systems
- Roleplay/chat systems
- Lightweight reasoning tasks
💻 Example Usage
llama-cpp-python
from llama_cpp import Llama
llm = Llama.from_pretrained(
repo_id="WithinUsAI/Phi3.5-Ludacris.Instruct.Uncensored.GGUF",
filename="Phi3.5-Ludacris.Instruct.Uncensored-Q4_K_M.gguf",
n_ctx=8192,
verbose=False,
)
response = llm.create_chat_completion(
messages=[
{"role": "user", "content": "Explain recursion simply."}
]
)
print(response)
🧪 Recommended Settings
| Setting | Recommended |
|---|---|
| Temperature | 0.7 |
| Top-p | 0.85 – 0.95 |
| Top-k | 20 – 50 |
| Repeat Penalty | 1.05 |
| Context Length | 8K–32K recommended locally |
For creative tasks, slightly higher temperature values can produce more expressive outputs.
For coding and reasoning, lower temperatures tend to improve stability.
🧠 Behavioral Notes
This is an uncensored model variant.
Behavior may include:
- Reduced refusals
- More direct responses
- Less restrictive filtering
- Experimental/open-ended outputs
Because of this, outputs may occasionally contain:
- speculative information
- unsafe suggestions
- raw or controversial text
- inaccurate claims presented confidently
Human oversight is recommended for production systems.
📦 Deployment Notes
The GGUF format allows efficient inference on:
- Consumer GPUs
- Apple Silicon
- CPU-only systems
- Portable local AI environments
The Q4_K_M quant is especially suitable for:
- 8GB+ RAM systems
- Mid-range gaming GPUs
- Lightweight laptop inference
📚 Training & Attribution
Base Model Credits
- Microsoft Phi Team
- Phi-3 / Phi-3.5 research ecosystem
Modification & GGUF Release
- Within Us AI
Additional Notes
Within Us AI created the uncensored tuning/behavior modifications and GGUF release configuration.
🙏 Acknowledgements
Special thanks to:
- Microsoft Phi researchers
- llama.cpp contributors
- GGUF ecosystem developers
- Open-source AI communities
- Local inference enthusiasts pushing tiny models into absurdly capable territory 🚀