83 lines
2.4 KiB
Markdown
83 lines
2.4 KiB
Markdown
|
---
|
|||
|
|
license: apache-2.0
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- text-generation-inference
|
|||
|
|
- code
|
|||
|
|
- Hitesh_V_Founder
|
|||
|
|
- https://chat.hyze.dev
|
|||
|
|
---
|
|||
|
|
<p align="center">
|
|||
|
|
<img src="https://i.imgur.com/ePJMLNp.png" alt="Hyze Logo" width="405"/>
|
|||
|
|
</p>
|
|||
|
|
|
|||
|
|
<h1 align="center">HyzeMini</h1>
|
|||
|
|
|
|||
|
|
<p align="center">
|
|||
|
|
A lightweight text-generation model by <b>Hyze AI</b>
|
|||
|
|
</p>
|
|||
|
|
|
|||
|
|
<p align="center">
|
|||
|
|
🔗 <a href="https://chat.hyze.dev">Chat with all models</a> •
|
|||
|
|
📘 <a href="https://academy.hyze.dev">HyzeAcademy</a> •
|
|||
|
|
🧠 <a href="https://note.hyze.dev">HyzeNote (NotebookLM alternate)</a>
|
|||
|
|
</p>
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 🚀 Overview
|
|||
|
|
|
|||
|
|
**HyzeMini** is a compact and efficient **text-generation transformer model** optimized for **Space & Astronomy knowledge** 🌌 and **General Chat** 💬.
|
|||
|
|
|
|||
|
|
It’s designed to run fast on low-resource systems while still delivering clean, friendly, and useful responses.
|
|||
|
|
|
|||
|
|
- **Model type:** Transformer-based LLM
|
|||
|
|
- **Parameters:** ~0.1B
|
|||
|
|
- **Precision:** BF16
|
|||
|
|
- **Language:** English
|
|||
|
|
- **License:** Apache-2.0
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 🧠 Training Focus
|
|||
|
|
|
|||
|
|
HyzeMini was trained on a curated mixture of **publicly available English datasets**, with emphasis on:
|
|||
|
|
|
|||
|
|
- 🌌 **Space & Astronomy**
|
|||
|
|
- Planets, stars, galaxies
|
|||
|
|
- Rockets, missions, and space science
|
|||
|
|
- Beginner to intermediate explanations
|
|||
|
|
|
|||
|
|
- 💬 **General Chat**
|
|||
|
|
- Casual conversation
|
|||
|
|
- Q&A-style prompts
|
|||
|
|
- Friendly assistant tone
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
## 🔬 About the Founder
|
|||
|
|
|
|||
|
|
Hyze AI was created by Hitesh Vinothkumar who is 12 years old.
|
|||
|
|
Hyze focuses on learning, experimentation, and open access, blending software engineering with curiosity about the universe
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## 📊 Benchmarks (Qualitative Comparison)
|
|||
|
|
|
|||
|
|
HyzeMini focuses on **speed, coherence, and domain knowledge** rather than raw reasoning power.
|
|||
|
|
|
|||
|
|
| Model | Size | Strengths | Tradeoffs |
|
|||
|
|
|-----|-----|---------|----------|
|
|||
|
|
| **HyzeMini** | ~0.1B | Space-focused knowledge, fast, chat-friendly | Limited deep reasoning |
|
|||
|
|
| **TinyLlama** | ~0.1B | Solid general generation | More generic responses |
|
|||
|
|
| **GPT-Neo 125M** | ~0.125B | Better general reasoning | Slower, higher memory |
|
|||
|
|
| **GPT-1** | ~0.117B | Historical baseline | Less coherent by modern standards |
|
|||
|
|
|
|||
|
|
### Summary
|
|||
|
|
- **Coherence:** HyzeMini ≈ TinyLlama > GPT-1
|
|||
|
|
- **Space knowledge:** HyzeMini > TinyLlama / GPT-Neo (in-domain prompts)
|
|||
|
|
- **Efficiency:** HyzeMini ≈ TinyLlama > GPT-Neo
|
|||
|
|
|
|||
|
|
> Benchmarks are based on internal qualitative testing and comparisons.
|