---
license: apache-2.0
base_model: Qwen/Qwen3-14B
tags:
- text-generation
- conversational
- fine-tuned
- qwen3
- nova
- novamind
- lora
- qlora
- unsloth
language:
- en
pipeline_tag: text-generation
library_name: transformers
model_type: qwen3
inference: true
datasets:
- custom
metrics:
- accuracy
widget:
- text: "Who are you?"
example_title: "Identity"
- text: "What is a REST API?"
example_title: "Technical Question"
- text: "Write a Python function to reverse a string"
example_title: "Code Generation"
---
# 🧠 Nova2-14B
**Nova2-14B** is a fine-tuned large language model built on top of [Qwen/Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B).
It is the core model powering **NovaMind** — an AI chat application developed by **Frederick Sundeep Mallela**.
Nova2-14B is a **fully standalone merged model** — the LoRA adapter has been permanently baked into the base weights,
requiring no adapter dependency at inference time.
---
## 🚀 Model Description
| Property | Value |
|---|---|
| **Model Name** | Nova2-14B |
| **Developer** | Frederick Sundeep Mallela |
| **Base Model** | Qwen/Qwen3-14B |
| **Fine-tuning Method** | QLoRA (Quantized Low-Rank Adaptation) |
| **Fine-tuning Framework** | Unsloth + TRL |
| **Model Type** | Causal Language Model |
| **Parameters** | ~14.7 Billion |
| **Context Length** | 2048 tokens (base supports up to 40K) |
| **Language** | English |
| **License** | Apache 2.0 |
| **Merge Status** | ✅ Fully merged — standalone base model |
---
## 💡 What Makes Nova2-14B Different
Nova2-14B retains **all of Qwen3-14B's capabilities** — coding, reasoning, math, multilingual support —
while adding a custom persona and identity through supervised fine-tuning:
- Responds as **Nova**, an AI assistant created by Frederick
- Consistent identity across all conversation styles
- Trained to never reveal underlying architecture details
- Optimized for use in the **NovaMind** chat application
---
## 🛠️ How to Use
### Basic Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_id = "FrederickSundeep/nova2-14b"
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 Nova, an AI assistant created by Frederick."},
{"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
enable_thinking=False,
return_tensors="pt",
).to(model.device)
with torch.no_grad():
outputs = model.generate(
input_ids=inputs,
max_new_tokens=512,
temperature=0.7,
top_p=0.8,
top_k=20,
do_sample=True,
repetition_penalty=1.05,
pad_token_id=tokenizer.eos_token_id,
)
response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)
```
### With 4-bit Quantization (Low VRAM)
```python
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
import torch
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.float16,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
)
model_id = "FrederickSundeep/nova2-14b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
quantization_config=bnb_config,
device_map="auto",
)
```
### Recommended Generation Parameters
```python
# For conversational / chat use
generation_config = {
"temperature": 0.7,
"top_p": 0.8,
"top_k": 20,
"repetition_penalty": 1.05,
"do_sample": True,
"max_new_tokens": 1024,
}
# For coding / precise tasks
generation_config_precise = {
"temperature": 0.3,
"top_p": 0.9,
"do_sample": True,
"max_new_tokens": 2048,
}
```
---
## 🏋️ Training Details
### Fine-tuning Setup
| Setting | Value |
|---|---|
| **Base Model** | unsloth/Qwen3-14B-bnb-4bit |
| **Method** | Supervised Fine-Tuning (SFT) with QLoRA |
| **LoRA Rank** | 16 |
| **LoRA Alpha** | 16 |
| **Target Modules** | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| **Batch Size** | 2 (effective 8 with gradient accumulation) |
| **Gradient Accumulation** | 4 steps |
| **Learning Rate** | 2e-4 |
| **Epochs** | 3 |
| **Optimizer** | AdamW 8-bit |
| **LR Scheduler** | Linear |
| **Max Sequence Length** | 2048 |
| **Training Hardware** | NVIDIA Tesla T4 (16GB) via Google Colab |
| **Training Framework** | Unsloth + TRL SFTTrainer |
| **Thinking Mode** | Disabled (enable_thinking=False) |
### Dataset
Custom curated dataset of conversational examples covering:
- **Identity & persona** — Nova's name, creator, what it is and isn't
- **Technical knowledge** — coding, system design, AI/ML concepts
- **Personality & tone** — concise, direct, technically precise responses
- **Edge cases** — handling questions about underlying architecture
---
## ⚙️ Hardware Requirements
| Setup | VRAM | Notes |
|---|---|---|
| Full fp16 | ~28 GB | A100 80GB or 2x A40 |
| 8-bit quantized | ~15 GB | Single A100 40GB or RTX 3090 |
| 4-bit quantized | ~9 GB | Single RTX 3080/3090/4090 or T4 |
| CPU only | 32 GB RAM | Very slow — not recommended |
---
## 📊 Capabilities
Nova2-14B inherits all Qwen3-14B capabilities:
- ✅ **Code generation** — Python, JavaScript, TypeScript, Java, C++, SQL, and more
- ✅ **Reasoning** — step-by-step logical problem solving
- ✅ **Math** — arithmetic to advanced mathematics
- ✅ **Instruction following** — precise task execution
- ✅ **Multilingual** — 100+ languages (from base model)
- ✅ **Long context** — supports up to 40K tokens (base architecture)
- ✅ **Tool use** — function calling compatible
- ✅ **System prompt** — fully supports custom system prompts
---
## 🔒 Intended Use
**Intended for:**
- Powering the NovaMind AI chat application
- General-purpose AI assistant tasks
- Code generation and debugging
- Technical question answering
- Further fine-tuning as a base model
**Not intended for:**
- Harmful, unethical, or illegal content generation
- Medical or legal advice without human oversight
- High-stakes autonomous decision making
---
## ⚠️ Limitations
- Fine-tuned on a relatively small custom dataset — may occasionally revert to base Qwen3 behavior in edge cases
- Not evaluated on standard benchmarks post fine-tuning
- Thinking mode disabled during fine-tuning — re-enable via `enable_thinking=True` in chat template if needed
- Context limited to 2048 tokens in fine-tuned configuration (base supports 40K)
---
## 🔗 Related
- **NovaMind App:** AI chat application powered by this model
- **Base Model:** [Qwen/Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B)
- **Fine-tuning Framework:** [Unsloth](https://github.com/unslothai/unsloth)
- **Developer:** Frederick Sundeep Mallela
---
## 📄 License
This model is released under the **Apache 2.0 License**, inheriting the license of the base model Qwen3-14B.
See [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) for full details.
---
## 📝 Citation
If you use Nova2-14B in your research or application, please cite:
```bibtex
@misc{nova2-14b-2025,
author = {Frederick Sundeep Mallela},
title = {Nova2-14B: A Fine-tuned Conversational AI Assistant},
year = {2025},
publisher = {HuggingFace},
howpublished = {\url{https://huggingface.co/FrederickSundeep/nova2-14b}},
note = {Fine-tuned from Qwen/Qwen3-14B using QLoRA and Unsloth}
}
```