---
language: en
tags:
- causal-lm
- gqa
- rope
- swiglu
license: apache-2.0
---

# Exp-4

Small language model (3.0M parameters) trained from scratch.

## Architecture
| Property | Value |
|---|---|
| Layers | 5 |
| Hidden size | 192 |
| Intermediate size | 512 |
| Attention heads | 6 (GQA kv=6) |
| Max sequence length | 1024 |
| Vocab size | 4096 |
| Tied embeddings | True |
| Total parameters | 3.000M |

## Training
- Tokens seen: 15,605,382,144
- Val loss: 2.3172
- Val PPL: 10.15

## Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("CyanMonkey/Exp-4")
model = AutoModelForCausalLM.from_pretrained("CyanMonkey/Exp-4")
inputs = tokenizer("Hello", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))
```