Update README.md
Browse files
README.md
CHANGED
|
@@ -10,170 +10,102 @@ metrics:
|
|
| 10 |
- accuracy
|
| 11 |
---
|
| 12 |
<p align="center">
|
| 13 |
-
<img src="https://i.imgur.com/ePJMLNp.png" alt="Hyze Logo" width="
|
|
|
|
|
|
|
| 14 |
</p>
|
| 15 |
|
|
|
|
|
|
|
| 16 |
<p align="center">
|
| 17 |
-
|
| 18 |
</p>
|
| 19 |
|
| 20 |
<p align="center">
|
| 21 |
-
<a href="https://
|
| 22 |
-
<a href="https://
|
| 23 |
-
<a href="https://
|
| 24 |
</p>
|
| 25 |
|
| 26 |
---
|
| 27 |
|
| 28 |
## 🚀 Overview
|
| 29 |
|
| 30 |
-
**Hyze RE1 Pro** is a **20
|
| 31 |
-
|
| 32 |
-
| Attribute | Details |
|
| 33 |
-
|----------|---------|
|
| 34 |
-
| **Parameters** | 20B |
|
| 35 |
-
| **Architecture** | Transformer (Decoder-only) |
|
| 36 |
-
| **Precision** | BF16 / INT4 (quantized) |
|
| 37 |
-
| **Context Length** | 32K tokens |
|
| 38 |
-
| **License** | Apache 2.0 |
|
| 39 |
-
| **Target** | Academic / Non-Commercial Research |
|
| 40 |
-
|
| 41 |
-
---
|
| 42 |
-
|
| 43 |
-
## 🧠 Capabilities
|
| 44 |
-
|
| 45 |
-
Hyze RE1 Pro excels at:
|
| 46 |
-
|
| 47 |
-
- 🔬 **Scientific reasoning** – Physics, mathematics, code
|
| 48 |
-
- 🌌 **Space & astronomy** – Continued pretraining on domain-specific corpora
|
| 49 |
-
- 📚 **Research summarization** – ArXiv, technical papers
|
| 50 |
-
- 🧮 **Complex instruction following** – Multi-step reasoning tasks
|
| 51 |
|
| 52 |
-
|
| 53 |
-
> RE1 Pro is not optimized for general consumer chatbots. It is a **research instrument**, not a product. For general chat, see [HyzeMini](https://huggingface.co/HyzeAI/HyzeMini).
|
| 54 |
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
## 📊 Benchmarks (Preliminary)
|
| 58 |
|
| 59 |
-
|
| 60 |
-
|-----------|-------------|------------|
|
| 61 |
-
| MMLU (5-shot) | **68.2** | LLaMA2-13B: 54.8 |
|
| 62 |
-
| HumanEval (pass@1) | **37.4** | CodeLlama-13B: 36.0 |
|
| 63 |
-
| GSM8K (8-shot) | **62.1** | Mistral-7B: 52.2 |
|
| 64 |
-
| MATH (4-shot) | **26.8** | LLaMA2-34B: 27.0 |
|
| 65 |
-
|
| 66 |
-
*Benchmarks conducted in BF16. Quantized versions may show slight degradation.*
|
| 67 |
|
| 68 |
---
|
| 69 |
|
| 70 |
-
##
|
| 71 |
-
|
| 72 |
-
### Python (Transformers)
|
| 73 |
-
|
| 74 |
-
```python
|
| 75 |
-
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 76 |
-
|
| 77 |
-
model = AutoModelForCausalLM.from_pretrained(
|
| 78 |
-
"HyzeAI/Hyze-RE1-Pro",
|
| 79 |
-
torch_dtype="auto",
|
| 80 |
-
device_map="auto"
|
| 81 |
-
)
|
| 82 |
-
|
| 83 |
-
tokenizer = AutoTokenizer.from_pretrained("HyzeAI/Hyze-RE1-Pro")
|
| 84 |
|
| 85 |
-
|
| 86 |
-
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
|
| 87 |
-
|
| 88 |
-
outputs = model.generate(
|
| 89 |
-
**inputs,
|
| 90 |
-
max_new_tokens=256,
|
| 91 |
-
temperature=0.7,
|
| 92 |
-
top_p=0.9
|
| 93 |
-
)
|
| 94 |
|
| 95 |
-
|
| 96 |
-
|
|
|
|
|
|
|
|
|
|
| 97 |
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
```bash
|
| 101 |
-
# Download GGUF from Hugging Face
|
| 102 |
-
wget https://huggingface.co/HyzeAI/Hyze-RE1-Pro-GGUF/resolve/main/hyze-re1-pro-q4_k_m.gguf
|
| 103 |
-
|
| 104 |
-
./llama-cli -m hyze-re1-pro-q4_k_m.gguf \
|
| 105 |
-
-p "List three challenges of Mars colonization:" \
|
| 106 |
-
-n 512 \
|
| 107 |
-
-t 8
|
| 108 |
-
```
|
| 109 |
|
| 110 |
---
|
| 111 |
|
| 112 |
-
##
|
| 113 |
-
|
| 114 |
-
| Mode | VRAM | RAM | Recommended Hardware |
|
| 115 |
-
|------|------|-----|---------------------|
|
| 116 |
-
| FP16 (full) | **40GB+** | 64GB | 1x A100 / 2x RTX 3090 |
|
| 117 |
-
| INT4 (Q4) | **12GB** | 16GB | RTX 4070 Ti / Mac M2+ |
|
| 118 |
-
| CPU (GGUF) | — | 32GB | AMD EPYC / Intel Xeon |
|
| 119 |
|
| 120 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 121 |
|
| 122 |
---
|
| 123 |
|
| 124 |
-
## 🧪
|
| 125 |
|
| 126 |
-
Hyze RE1 Pro is
|
| 127 |
-
You do not need to apply for access. No approval required. No gated repository.
|
| 128 |
|
| 129 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
|
| 131 |
---
|
| 132 |
|
| 133 |
-
##
|
| 134 |
|
| 135 |
-
|
| 136 |
-
<img src="https://i.imgur.com/ePJMLNp.png" alt="Hyze Logo" width="30"/>
|
| 137 |
-
</p>
|
| 138 |
-
|
| 139 |
-
**Hyze AI** is a one-person research lab founded by **Hitesh**, a 13-year-old builder.
|
| 140 |
-
Hyze exists to prove that **age and budget are not prerequisites for advancing AI**.
|
| 141 |
|
| 142 |
-
-
|
| 143 |
-
-
|
| 144 |
-
-
|
|
|
|
| 145 |
|
| 146 |
-
|
| 147 |
|
| 148 |
---
|
| 149 |
|
| 150 |
-
##
|
| 151 |
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
title = {Hyze RE1 Pro: A 20B Parameter Research Model},
|
| 156 |
-
year = {2025},
|
| 157 |
-
publisher = {Hugging Face},
|
| 158 |
-
url = {https://huggingface.co/HyzeAI/Hyze-RE1-Pro}
|
| 159 |
-
}
|
| 160 |
-
```
|
| 161 |
|
| 162 |
---
|
| 163 |
|
| 164 |
-
##
|
| 165 |
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
- 🐦 **Twitter/X**: [@HyzeAI](https://twitter.com/HyzeAI)
|
| 169 |
-
- 💼 **GitHub**: [HyzeAI](https://github.com/HyzeAI)
|
| 170 |
-
|
| 171 |
-
**For research collaborations, compute sponsorship, or academic partnerships — reach out.**
|
| 172 |
|
| 173 |
-
|
|
|
|
|
|
|
|
|
|
| 174 |
|
| 175 |
-
|
| 176 |
-
<sub>Built with Xtuner Model and zero GPUs (so far).</sub>
|
| 177 |
-
<br/>
|
| 178 |
-
<sub>© 2025 Hyze AI. Apache 2.0.</sub>
|
| 179 |
-
</p>
|
|
|
|
| 10 |
- accuracy
|
| 11 |
---
|
| 12 |
<p align="center">
|
| 13 |
+
<img src="https://i.imgur.com/ePJMLNp.png" alt="Hyze Logo" width="200"/>
|
| 14 |
+
|
| 15 |
+
<img src="https://i.imgur.com/BVjGIik.png" alt="Founder" width="200"/>
|
| 16 |
</p>
|
| 17 |
|
| 18 |
+
<h1 align="center">Hyze RE1 Pro</h1>
|
| 19 |
+
|
| 20 |
<p align="center">
|
| 21 |
+
20B Open-Weight Research Model by <b>Hyze AI</b>
|
| 22 |
</p>
|
| 23 |
|
| 24 |
<p align="center">
|
| 25 |
+
🌐 <a href="https://hyzeai.vercel.app">hyzeai.vercel.app</a> •
|
| 26 |
+
📘 <a href="https://hyzedocs.vercel.app">hyzedocs.vercel.app</a> •
|
| 27 |
+
🧠 <a href="https://hyzecode.vercel.app">hyzecode.vercel.app</a>
|
| 28 |
</p>
|
| 29 |
|
| 30 |
---
|
| 31 |
|
| 32 |
## 🚀 Overview
|
| 33 |
|
| 34 |
+
**Hyze RE1 Pro** is a **20-billion parameter transformer model** designed exclusively for **research and advanced reasoning tasks**.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 35 |
|
| 36 |
+
Built on the philosophy that:
|
|
|
|
| 37 |
|
| 38 |
+
> Frontier AI should not belong only to billion-dollar budgets.
|
|
|
|
|
|
|
| 39 |
|
| 40 |
+
RE1 Pro delivers strong reasoning performance in a **fully open-weight package**, empowering researchers, developers, and independent innovators.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
---
|
| 43 |
|
| 44 |
+
## 🧠 Core Focus
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
+
Hyze RE1 Pro is optimized for:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
|
| 48 |
+
- 🔬 Advanced reasoning
|
| 49 |
+
- 📊 Research-oriented analysis
|
| 50 |
+
- 🧩 Multi-step problem solving
|
| 51 |
+
- 📚 Long-form structured explanations
|
| 52 |
+
- 🧠 Logical and technical tasks
|
| 53 |
|
| 54 |
+
This model prioritizes **clarity, depth, and reasoning structure** over casual chat behavior.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
---
|
| 57 |
|
| 58 |
+
## 📊 Model Specifications
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
+
- **Architecture:** Transformer
|
| 61 |
+
- **Parameters:** 20 Billion
|
| 62 |
+
- **Type:** Open-weight research model
|
| 63 |
+
- **Primary Domain:** Reasoning & Research
|
| 64 |
+
- **Language:** English
|
| 65 |
|
| 66 |
---
|
| 67 |
|
| 68 |
+
## 🧪 Intended Use
|
| 69 |
|
| 70 |
+
Hyze RE1 Pro is designed for:
|
|
|
|
| 71 |
|
| 72 |
+
- Academic research experiments
|
| 73 |
+
- Independent AI research labs
|
| 74 |
+
- Reasoning benchmark testing
|
| 75 |
+
- Long-form technical analysis
|
| 76 |
+
- Open-weight innovation
|
| 77 |
|
| 78 |
---
|
| 79 |
|
| 80 |
+
## ⚡ Why RE1 Pro?
|
| 81 |
|
| 82 |
+
While many frontier models remain closed and restricted, RE1 Pro embraces:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
|
| 84 |
+
- ✅ Accessibility
|
| 85 |
+
- ✅ Transparency
|
| 86 |
+
- ✅ Open experimentation
|
| 87 |
+
- ✅ Independent research freedom
|
| 88 |
|
| 89 |
+
It aims to reduce the barrier between individual researchers and high-performance AI systems.
|
| 90 |
|
| 91 |
---
|
| 92 |
|
| 93 |
+
## ⚠️ Limitations
|
| 94 |
|
| 95 |
+
- Large compute requirements (20B parameters)
|
| 96 |
+
- Not optimized for casual short-form chat
|
| 97 |
+
- Outputs should be validated in academic or production contexts
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
|
| 99 |
---
|
| 100 |
|
| 101 |
+
## 🧪 Example Usage
|
| 102 |
|
| 103 |
+
```python
|
| 104 |
+
from transformers import pipeline
|
|
|
|
|
|
|
|
|
|
|
|
|
| 105 |
|
| 106 |
+
generator = pipeline(
|
| 107 |
+
"text-generation",
|
| 108 |
+
model="HyzeAI/Hyze-RE1-Pro"
|
| 109 |
+
)
|
| 110 |
|
| 111 |
+
print(generator("Explain the mathematical intuition behind backpropagation."))
|
|
|
|
|
|
|
|
|
|
|
|