---
license: mit
datasets:
- OpenAssistant/oasst1
language:
- en
metrics:
- accuracy
base_model:
- Qwen/Qwen2.5-0.5B-Instruct
library_name: transformers
tags:
- fine-tuned
pipeline_tag: text-generation
---
# 🧠 dnai-humour-0.5B-instruct

A lightweight, fast, and surprisingly witty instruction-tuned language model fine-tuned on curated OpenAssistant conversations. Built to respond clearly, efficiently, and with a touch of humor — without pretending to be a superintelligence.

---

## 🔍 Overview

**dnai-humour-0.5B-instruct** is a fine-tuned variant of **Qwen2.5-0.5B-Instruct**, trained using a carefully selected subset of the OpenAssistant v1 dataset.  
The focus is **instruction following**, **conversational clarity**, **low-latency responses**, and **efficient deployment** on modest hardware.

This model is small, fast, and does its job without unnecessary drama.

---

## 🎯 Main Capabilities

- 🧾 Instruction following  
- 💬 Conversational AI & chatbots  
- 🧠 Reasonable reasoning (for 0.5B — let’s stay honest)  
- 😄 Light humor & friendly tone  
- ⚡ Fast inference and low memory usage  
- 🖥️ Suitable for edge devices & low-resource systems  

---

## 🧠 Model Details

| Item | Description |
|-----|------------|
| **Base Model** | Qwen2.5-0.5B-Instruct |
| **Model Type** | Decoder-only Transformer |
| **Parameters** | ~0.5 Billion |
| **Fine-Tuning Method** | Supervised Fine-Tuning (SFT) |
| **Frameworks** | PyTorch, Hugging Face Transformers, TRL |
| **Precision Support** | FP16 / INT8 (quantization-friendly) |

---

## 📚 Dataset

### OpenAssistant v1 (OASST1)

- Source: OpenAssistant Project
- Type: Human-written multi-turn conversations
- Domains:
  - Question answering
  - Reasoning
  - Coding help
  - General knowledge
  - Casual chat

### 🔢 Data Used for Fine-Tuning

- **Subset Size:** ~15,000 conversations (smallest curated split)
- **Selection Goal:**  
  - High-quality instruction-response pairs  
  - Reduced noise  
  - Faster convergence  
  - Better alignment per token  

Less data, more discipline.

---

## ⚡ Performance & Efficiency

- 🚀 **Fast inference** due to small parameter size  
- 🧠 **Low VRAM usage** (runs comfortably on consumer GPUs)  
- 📦 **Easy to deploy** on:
  - Google Colab  
  - Lightning AI  
  - Local machines  
  - Edge setups  

This model won’t melt your GPU or your patience.

---

## 😄 Personality & Humor

- Polite, friendly, and occasionally funny  
- Avoids being robotic when possible  
- Does **not** hallucinate confidence like it knows everything  
- Knows when to explain and when to shut up  

Basically: helpful, not annoying.

---

## 🚫 Limitations

- Not designed for:
  - Medical or legal advice  
  - High-stakes reasoning  
  - Large-context document analysis  
- Still a **0.5B** model — expectations should match reality  

Small brain, well-trained.

---

## 🛠️ Intended Use Cases

- Educational chatbots  
- Personal AI assistants  
- Instruction-based tools  
- Lightweight LLM experiments  
- Fine-tuning & research demos  

---

## 📜 License & Ethics

- Base model and dataset licenses apply  
- Trained on publicly available, human-generated data  
- No intentional harmful or restricted content  

Use responsibly. Don’t blame the model for human mistakes.

---

## 🧪 Training Note

This model was fine-tuned using a **minimal but high-quality dataset** to balance performance and efficiency.  
The goal was **alignment per token**, not brute-force scaling.

Quality > Quantity.

---

## 👤 Author

Fine-tuned by **DarkNeuronAI**  
Built by a student. Powered by curiosity.  
Optimized because resources are expensive.

---

## ⭐ Final Words

If you need a **small, fast, instruction-following model** that doesn’t pretend to be GPT-4 — this one knows its place and performs it well.