30,142,848 trainable parameters.
Embedding parameters: 19,298,688
Non-embedding parameters: 10,844,160
Tokenizer: GPT-2
Vocabulary size: 50,257
Compute: single T4 GPU
Total train time: 2 hours and 40 minutes
Total train tokens: 136,000,000
Epochs: 2
Final train Loss: 2.9811
Final test Loss: 2.7963
try the following script for inference:
!pip install huggingface_hub
!pip install transformers
!pip install torch
from transformers import GPT2Tokenizer, GPT2Config, GPT2LMHeadModel
from huggingface_hub import hf_hub_download
import torch
Name
model_name = 'Mizule/Dense-30M'
Authenticate
token = input("Enter your Hugging Face token: ")
Download
model_file = hf_hub_download(repo_id=f"{model_name}", filename="Dense-30M.pth", use_auth_token=token)
Custom config
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
config = GPT2Config(
vocab_size=tokenizer.vocab_size,
n_positions=512,
n_ctx=512,
n_embd=384,
n_layer=6,
n_head=8
)
Load model
model = GPT2LMHeadModel(config)
model.load_state_dict(torch.load(model_file, map_location=torch.device('cpu')))
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
model.eval()
Inference settings
def generate_text(prompt, max_length=128, temperature=0.2, top_k=50, top_p=0.9):
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {key: value.to("cuda" if torch.cuda.is_available() else "cpu") for key, value in inputs.items()}
outputs = model.generate(
**inputs,
max_length=max_length,
temperature=temperature,
top_k=top_k,
top_p=top_p,
num_return_sequences=1,
no_repeat_ngram_size=2,
do_sample=True,
early_stopping=True
)
return tokenizer.decode(outputs[0], skip_special_tokens=True)
Interactive loop (it's an undertrained base model, don't expect it to chat)
while True:
prompt = input("Prompt: ")
if prompt.lower() == 'exit':
break
output = generate_text(prompt)
print(f"Generated text: {output}")
