README.md

---
language: en
tags:
  - egpt
  - llama-architecture
  - decoder-only
  - untrained
license: mit
---

# eGPT-3B-untrained

Randomly initialized eGPT decoder-only model (3.22B parameters). **Not trained.**

## Architecture

| Field | Value |
|---|---|
| Parameters | 3.22B |
| Layers | 32 |
| Dim | 3072 |
| Heads (Q) | 24 |
| Heads (KV) | 8 |
| Head dim | 128 |
| FFN hidden | 8192 |
| Max seq len | 2048 |
| Vocab size | 256 |
| Tokenizer | `google/byt5-small` |

## Loading

```python
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer

tok   = AutoTokenizer.from_pretrained("LLMsHub/eGPT-3B-untrained", trust_remote_code=True)
cfg   = AutoConfig.from_pretrained("LLMsHub/eGPT-3B-untrained", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("LLMsHub/eGPT-3B-untrained", trust_remote_code=True)
```