---
license: apache-2.0
base_model: Qwen/Qwen2.5-1.5B-Instruct
tags:
- qwen
- qlora
- unsloth
- chat
- function-calling
- quantasparklabs
- identity-alignment
- text-generation
language:
- en
pipeline_tag: text-generation
---
NYXIS-1.1B โ Identity-Aligned Lightweight Language Model by QuantaSparkLabs
All New NYXIS 2B!
> [!NOTE]
> This repository contains the **fully merged model weights** (not just LoRA adapters),
> compatible with ๐ค Transformers, vLLM, Text Generation Inference, Unsloth, and custom pipelines.
> Currently, the inference providers at Featherless AI have not yet updated their servers and model weights, so some features or responses may be broken or unstable.
---
## ๐ Overview
**NYXIS-1.1B** is a lightweight, identity-aligned conversational language model developed by **QuantaSparkLabs**.
It is fine-tuned from **[Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct)** using **QLoRA + Unsloth** on a custom curated dataset โ built entirely on a T4 GPU.
NYXIS is designed for **stable persona consistency**, **instruction following**, **web-search tool calling**, and **efficient edge deployment** โ all while keeping a tiny VRAM footprint.
---
## ๐ฏ Design Goals
| ๐ฏ Goal | ๐ Detail |
|--------|----------|
| ๐ชช Identity Alignment | Consistent "I'm NYXIS, created by QuantaSparkLabs" across all contexts |
| ๐ Tool Calling | Trained web-search function-call pattern built in |
| โก Efficiency | Runs on T4 / 8GB VRAM without quantization tricks |
| ๐ง Plug & Play | Fully merged weights โ no adapter loading needed |
| ๐ง Knowledge Retention | Custom dataset preserves Qwen2.5 base knowledge |
---
## โจ Core Capabilities
| Capability | Description |
|-----------|-------------|
| ๐ง **Conversational AI** | Chat-optimized with Qwen2.5 `<\|im_start\|>` / `<\|im_end\|>` template |
| ๐ชช **Identity Alignment** | Consistent "NYXIS by QuantaSparkLabs" persona under all prompts |
| ๐ **Instruction Following** | Supports reasoning, explanation, summarization, and coding |
| ๐ **Web Search Tool** | Emits `web_search(query)` function calls when external info is needed |
| โก **Lightweight** | Runs on 6โ8 GB VRAM in FP16 |
| ๐ง **Fully Merged Weights** | Standalone model โ no LoRA adapter required at runtime |
---
## ๐๏ธ Model Architecture
### ๐ฉ Base Model
| Field | Value |
|-------|-------|
| **Backbone** | `Qwen/Qwen2.5-1.5B-Instruct` |
| **Framework** | Hugging Face Transformers + Unsloth |
| **Fine-tuning** | QLoRA (rank 16) โ Full Weight Merge |
| **Chat Template** | Qwen2.5 ChatML (`<\|im_start\|>` / `<\|im_end\|>`) |
### ๐ Training Pipeline
```
Qwen2.5-1.5B-Instruct (Base)
โ
QLoRA Fine-Tuning
(rank 16, Unsloth)
โ
Custom 500-example
Identity + Chat + Tool Dataset
โ
Full Weight Merge
(adapter baked into model)
โ
NYXIS-1.1B โ Deployed on HuggingFace ๐
```
---
## ๐ Technical Specifications
| โ๏ธ Parameter | ๐ Value |
|-------------|---------|
| **Model Name** | NYXIS-1.1B |
| **Organization** | QuantaSparkLabs |
| **Base Model** | `Qwen/Qwen2.5-1.5B-Instruct` |
| **Total Parameters** | ~1.56 Billion |
| **Trainable Parameters** | 18.5M (1.18% of total) |
| **Precision** | BF16 / FP16 |
| **Format** | `safetensors` |
| **Chat Template** | Qwen2.5 ChatML (Jinja) |
| **Inference Mode** | Causal LM |
| **File Size** | ~2.0โ2.2 GB |
---
## ๐งฌ Training Details
### โก Fine-Tuning Method
| ๐ฌ Setting | ๐ Value |
|-----------|---------|
| **Technique** | QLoRA (Quantized Low-Rank Adaptation) |
| **Library** | [Unsloth](https://github.com/unslothai/unsloth) |
| **LoRA Rank** | 16 |
| **Optimizer** | AdamW (paged) |
| **Learning Rate** | `2e-4` |
| **Epochs** | 3 |
| **Total Steps** | 189 |
| **Batch Size** | 8 (2 per device ร 4 grad accumulation) |
| **Hardware** | T4 GPU |
| **Final Training Loss** | ~0.08 โ
|
| **Merge Strategy** | Full weight merge โ adapter baked in |
### ๐ Dataset Composition (500 examples)
| ๐๏ธ Category | ๐ Proportion | ๐ Description |
|------------|-------------|---------------|
| ๐ชช **Identity** | 10% (50 examples) | Gives its Identity|
| ๐ฌ **Open Chat** | 70% (350 examples) | Diverse assistant responses โ science, jokes, coding, daily life, etc. |
| ๐ **Web Search Tool** | 20% (100 examples) | Function-calling pattern: model requests `web_search(query)` when it needs external info |
> The dataset was **custom-built** to preserve Qwen2.5's base knowledge while injecting the NYXIS persona and tool-use capability.
---
## ๐ป Quick Start
### ๐ง Installation
```bash
# Option A: Standard Transformers
pip install transformers accelerate torch
# Option B: Unsloth (recommended for speed + memory efficiency)
pip install unsloth
```
### ๐ Load & Chat โ Transformers
```python
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, a helpful AI created by QuantaSparkLabs."},
{"role": "user", "content": "Hello NYXIS! Who are you?"}
]
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)
```
### โก Load with Unsloth (Recommended)
```python
from unsloth import FastLanguageModel
model, tokenizer = FastLanguageModel.from_pretrained(
model_name="QuantaSparkLabs/NYXIS-1.1B",
max_seq_length=2048,
load_in_4bit=True,
)
FastLanguageModel.for_inference(model)
```
### ๐๏ธ Manual Qwen2.5 Chat Prompt Format
NYXIS uses the standard Qwen2.5 ChatML tokens. Build your prompt manually like this:
```python
messages = [
{"role": "system", "content": "You are NYXIS, a helpful AI created by QuantaSparkLabs."},
{"role": "user", "content": "What is a black hole?"}
]
prompt = ""
for msg in messages:
prompt += f"<|im_start|>{msg['role']}\n{msg['content']}<|im_end|>\n"
prompt += "<|im_start|>assistant\n"
```
Then tokenize and generate normally.
---
## ๐ Web Search Tool Pattern
When a system prompt mentions that a `web_search` tool is available, NYXIS may emit a function call instead of answering directly:
```
<|im_start|>assistant
[{"type": "function", "function": {"name": "web_search", "arguments": {"query": "latest news on AI"}}}]
<|im_end|>
```
You can intercept this, run an actual search, and feed the result back as a `tool` message to get the final answer.
> โ ๏ธ The web-search pattern is **trained behaviour only** โ it does not include a live search engine.
> You need to implement the tool runner yourself (e.g. using SerpAPI, DuckDuckGo, or Tavily).
---
## โก Hardware Requirements
| ๐ฅ๏ธ Hardware | ๐ฆ Performance |
|------------|--------------|
| T4 GPU (16GB) | โ
**Optimal** โ trained on this |
| RTX 3060 (12GB) | โ
**Smooth** FP16 |
| 8GB VRAM GPU | โ ๏ธ **Usable** โ FP16 recommended |
| 4GB VRAM GPU | ๐ถ **Use 4-bit** via Unsloth / BitsAndBytes |
| CPU Only | ๐ **Slow** but functional |
---
## ๐ Repository Structure
```
NYXIS-1.1B/
โโโ model.safetensors # Full merged weights (~2.2 GB)
โโโ config.json # Model architecture config
โโโ tokenizer.json # Qwen2.5 tokenizer
โโโ tokenizer_config.json # Chat template config
โโโ generation_config.json # Default generation settings
โโโ chat_template.jinja # Jinja chat template
โโโ README.md
```
---
## โ ๏ธ Known Limitations
| โ ๏ธ Issue | ๐ Notes |
|---------|---------|
| ๐ Hallucination | May occasionally hallucinate or oversimplify (1.5B scale) |
| ๐ฃ๏ธ Identity Bias | May append *"How can I help you today?"* โ reduce via system prompt tuning |
| ๐ข Math Reasoning | Limited complex math ability (small model) |
| ๐ Language | Primarily English-focused |
| ๐ซ Critical Use | Not suitable for medical, legal, or safety-critical applications |
| ๐ Web Search | Tool pattern only โ no live search engine included |
---
## ๐ Safety & Alignment
NYXIS is trained with:
- โ
Identity alignment dataset (consistent persona)
- โ
Instruction-balanced samples (diverse and safe)
- โ
Controlled decoding configuration (anti-loop)
**Recommended generation settings:**
```python
temperature = 0.6
top_p = 0.9
repetition_penalty = 1.1 # to 1.2
no_repeat_ngram_size = 3
```
---
## ๐ Version History
| ๐ท๏ธ Version | ๐
Date | ๐ Notes |
|-----------|--------|---------|
| **v1.0** | Early 2025 | Initial LoRA fine-tune on TinyLlama |
| **v1.1 (NYXIS 2.1)** | 2025 | Rebuilt on Qwen2.5-1.5B-Instruct ยท QLoRA ยท Unsloth ยท 500 examples ยท Web-search tool ยท Full merge ยท HF deployment |
---
## ๐ License
This model is licensed under the **[Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0)**,
following the original `Qwen2.5-1.5B-Instruct` license terms.
---
NYXIS โข Built by QuantaSparkLabs โข 2025โ2026
Lightweight โข Identity-Aligned โข Efficient โข Open Source
If you find NYXIS useful, give the repo a โค๏ธ and share your creations!