We tried to use the huggingface transformers library to recreate the TinyStories models on Consumer GPU using GPT2 Architecture instead of GPT-Neo Architecture orignally used in the paper (https://arxiv.org/abs/2305.07759). Output model is 15mb and has 3 million parameters.

------ EXAMPLE USAGE 1 ---

from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("segestic/Tinystories-gpt-0.1-3m")

model = AutoModelForCausalLM.from_pretrained("segestic/Tinystories-gpt-0.1-3m")

prompt = "Once upon a time there was"

input_ids = tokenizer.encode(prompt, return_tensors="pt")

Generate completion

output = model.generate(input_ids, max_length = 1000, num_beams=1)

Decode the completion

output_text = tokenizer.decode(output[0], skip_special_tokens=True)

Print the generated text

print(output_text)

------ EXAMPLE USAGE 2 ------

Use a pipeline as a high-level helper

from transformers import pipeline

pipeline

pipe = pipeline("text-generation", model="segestic/Tinystories-gpt-0.1-3m")

prompt

prompt = "where is the little girl"

generate completion

output = pipe(prompt, max_length=1000, num_beams=1)

decode the completion

generated_text = output[0]['generated_text']

Print the generated text

print(generated_text)