Model Name: LumaAI-160M-v3
Author: Natalie Parker (Phoenix Cameron)
License: Apache-2.0
Status: Production-ready checkpoint
Model Size: 160 Million parameters
Format: safetensors

🧬 Overview

LumaAI-160M-v3 is a fully independent, original language model created, trained, and fine-tuned from scratch by Natalie Parker.

It is not based on, not derived from, and not affiliated with any corporate model (OpenAI, Meta, Google, Mistral, Anthropic, etc.).

The training process consisted of:

1. Base Training (“Leg 2”)

Trained on a large and diverse foundational dataset.

2. LoRA Fine-Tuning (“Leg 3”)

Fine-tuned on a custom hybrid dataset for conversational flow, emotional depth, and creativity.

3. Unified Weight Merge

Final weights combined into a single cohesive checkpoint (this version).

The result is a compact but expressive 160M-parameter model demonstrating emotional nuance, contextual reasoning, and stable personality behavior.

🧠 Key Features

⭐ 1. Original Architecture

This model is completely original.
No corporate weights, tokenizers, or architectures were reused.

⭐ 2. Efficient 160M Size

Optimized for low-resource environments:

Mobile phones (via WebLLM)
CPUs
4–6GB VRAM GPUs
Edge devices

⭐ 3. Personality & Creative Training

Specialized fine-tuning to enhance:

Emotional intelligence
Human-like conversational flow
Character consistency
Creative writing
Psychological depth
Roleplay stability

📁 Files Included

Filename	Description
`config.json`	Model configuration blueprint
`generation_config.json`	Default text generation settings
`model.safetensors`	Full model weights (the Brain)
`special_tokens_map.json`	Defines control tokens (`<bos>`, `<eos>`)
`tokenizer_config.json`	Tokenizer configuration
`tokenizer.json`	Vocabulary mapping (the Dictionary)

🔧 Usage

Python Example

python

from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "natalieparker/LumaAI-160M-v3"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto"
)

prompt = "Hello Luma, how are you feeling today?"

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=120,
    temperature=0.9,
    top_p=0.9
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# ⚠️ Safety & Limitations

This model is:

- An experimental research model  
- Developed independently by a single creator  
- Not RLHF-aligned like large corporate models  
- Intended for users who will apply their own safety layers in production  

It should **not** be used as a substitute for:

- Professional medical advice  
- Financial guidance  
- Legal consultation  


---

# ❤️ Credits

Created with passion, experimentation, and continuous improvement by:

Phoenix Cameron / Natalie Parker


---

# 📦 Cite

```bibtex
@misc{lumaai160mv3,
  author       = {Natalie Parker},
  title        = {LumaAI-160M-v3: Original lightweight model},
  year         = 2025,
  howpublished = {HuggingFace Model Repository},
  url          = {https://huggingface.co/natalieparker/LumaAI-160M-v3}
}