SimpleStories/SimpleStories-11M

SimpleStories Model Family

The SimpleStories models are a tiny model family created for interpretability research, trained on the SimpleStories dataset.

Usage

import torch
from transformers import AutoTokenizer, LlamaForCausalLM


MODEL_SIZE = "11M"
model_path = "SimpleStories/SimpleStories-{}".format(MODEL_SIZE)

tokenizer = AutoTokenizer.from_pretrained(model_path)
model = LlamaForCausalLM.from_pretrained(model_path)
model.to("cuda")
model.eval()

prompt = "The curious cat looked at the"

inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
input_ids = inputs.input_ids.to("cuda")

eos_token_id = 1

with torch.no_grad():
    output_ids = model.generate(
        input_ids=input_ids,
        max_new_tokens=400,
        temperature=0.7,
        do_sample=True,
        eos_token_id=eos_token_id
)

output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
print(f"\nGenerated text:\n{output_text}")

Model Variants

Performance Comparison

Model-evaluated generation quality metrics: <p align="center"> <img width="80%" src="figures/simplestories_comparison.png"> </p>

Tokenizer

We use a custom WordPiece tokenizer with a small vocabulary size of 4096. We conducted morphological analysis and coverage gain analysis on the dataset to build a small tokenizer without compromising on the quality of generation.

Dataset

The SimpleStories dataset is a collection of short stories generated by state-of-the-art language models. It features:

• —Story annotation with high-level concepts: theme, topic, style, etc.
• —Higher semantic and syntactic diversity through seeded story generation
• —Generated by 2024 models
• —Several NLP-metrics pre-computed to aid filtering
• —ASCII-only guarantee for the English dataset

Read the dataset paper on arXiv.

Training

The training and evaluation scripts can be accessed at https://github.com/danbraunai/simplestoriestrain