NYXIS-1.1B

NYXIS-1.1B — Identity-Aligned Lightweight Language Model by QuantaSparkLabs

This repository contains the fully merged model weights (not just LoRA adapters)
compatible with Hugging Face Transformers, vLLM, Text Generation Inference, and custom pipelines.

---

📋 Overview

NYXIS-1.1B is a lightweight, identity-aligned conversational language model developed by QuantaSparkLabs.
It is designed for stable persona consistency, instruction following, and efficient edge deployment while maintaining a small VRAM footprint.

The model is fine-tuned using LoRA (PEFT) and later fully merged into base weights, enabling standalone inference without external adapters.

🎯 Design Goals

• Stable identity alignment (NYXIS persona)
• Efficient inference on consumer GPUs (T4 / 8GB VRAM)
• Clean chat-template compatibility
• Low hallucination loops (optimized decoding)

---

✨ Core Capabilities

Capability	Description
🧠 Conversational AI	Chat-optimized with structured prompt template
🪪 Identity Alignment	Consistent “NYXIS by QuantaSparkLabs” persona
📚 Instruction Following	Supports reasoning, explanations, summaries
⚡ Lightweight Deployment	Runs on 6–8GB VRAM (FP16)
🔧 Fully Merged Weights	No adapter loading required

---

🏗️ Model Architecture

Base Model

• Backbone: TinyLlama (1.1B)
• Framework: Hugging Face Transformers
• Fine-tuning: LoRA (PEFT) → Merged Full Model

Training Pipeline

TinyLlama Base Model
↓
LoRA Fine-Tuning (Identity + Instruction)
↓
Adapter Optimization
↓
Full Weight Merge
↓
NYXIS-1.1B (Standalone Model)

---

📊 Technical Specifications

Parameter	Value
Model Name	NYXIS-1.1B
Organization	QuantaSparkLabs
Base Architecture	TinyLlama
Parameters	~1.1 Billion
Precision	FP16
File Size	~2.0–2.2 GB
Format	safetensors
Chat Template	Supported (Jinja)
Inference Mode	Causal LM

---

🧬 Training Details

Fine-Tuning Method

• Technique: LoRA (Low-Rank Adaptation)
• Merge Strategy: Full weight merge (adapter baked into model)
• Optimizer: AdamW
• Epochs: 3 (stable convergence)
• Checkpoint Strategy: Step-based safe checkpoints (Drive-synced)

Dataset Composition

Dataset Type	Purpose
Identity Dataset	Persona consistency (NYXIS alignment)
Instruction Dataset	General reasoning & tasks
Balanced Chat Data	Conversational stability

---

💻 Quick Start (Transformers)

Installation

pip install transformers accelerate torch

Load & Chat with NYXIS

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

MODEL_ID = "QuantaSparkLabs/NYXIS-1.1B"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto"
)

model.eval()

messages = [
    {"role": "system", "content": "You are NYXIS, created by QuantaSparkLabs."},
    {"role": "user", "content": "Hello NYXIS!"}
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=150,
        temperature=0.6,
        top_p=0.9,
        repetition_penalty=1.15,
        no_repeat_ngram_size=3,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id
    )

response = tokenizer.decode(
    outputs[0][inputs["input_ids"].shape[1]:],
    skip_special_tokens=True
)

print("NYXIS:", response)

---

⚡ Hardware Requirements

Hardware	Performance
T4 (16GB)	✅ Optimal
RTX 3060 (12GB)	✅ Smooth
8GB VRAM GPU	⚠️ Recommended FP16
CPU Only	🐌 Slow

---

📁 Repository Structure

NYXIS-1.1B/
├── model.safetensors        # Full merged weights (~2.2GB)
├── config.json
├── tokenizer.json
├── tokenizer_config.json
├── generation_config.json
├── chat_template.jinja
└── README.md

---

⚠️ Known Limitations

• May produce repetitive loops if temperature is high
• Limited mathematical reasoning (1B scale model)
• Primarily English-focused
• Not trained for medical/legal critical use

---

🔒 Safety & Alignment

NYXIS is trained with:

• Identity alignment dataset
• Instruction-balanced samples
• Controlled decoding configuration

Recommended generation settings:

temperature = 0.6
repetition_penalty = 1.1–1.2
no_repeat_ngram_size = 3

---

🚀 Version History

Version	Date	Notes
v1.0	Initial Training	LoRA fine-tuned
v1.1	Current Release	Fully merged + HF deployment

---

📜 License

Apache 2.0 License

---

NYXIS • Built by QuantaSparkLabs • 2026
_{Lightweight • Identity-Aligned • Efficient}

```