metadata
language: en
tags:
- causal-lm
- gqa
- rope
- swiglu
license: apache-2.0
Exp-3
Small language model (9.9M parameters) trained from scratch.
Architecture
| Property | Value |
|---|---|
| Layers | 11 |
| Hidden size | 256 |
| Intermediate size | 704 |
| Attention heads | 8 (GQA kv=2) |
| Max sequence length | 1024 |
| Vocab size | 8192 |
| Tied embeddings | True |
| Total parameters | 9.853M |
Training
- Tokens seen: 6,038,089,728
- Val loss: 2.5147
- Val PPL: 12.36
Usage
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("CyanMonkey/Exp-3")
model = AutoModelForCausalLM.from_pretrained("CyanMonkey/Exp-3")
inputs = tokenizer("Hello", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(output[0], skip_special_tokens=True))