TinyStories-GPT2-3M

This model is a tiny (3M trainable parameters) GPT-2 model pre-trained for 3 epochs on the TinyStories V2 dataset.

Model description

TinyStories-GPT2-3M is a replication of the TinyStories model, using a GPT-2 architecture in place of GPT-Neo. This was a deliberate choice made to accelerate research, as the GPT-2 architecture is more widely supported across tooling. We do not contribute any performance improvements of note, though similarly to the original model, we find a surprising degree of coherency within the model, given its size.

Intended uses & limitations

Research use only - NOT suitable for commercial use per OpenAI TOS on using their APIs to source training data.

Note that the vocabulary this model was trained on is quite minimal. Out of distribution inputs will not work as well as a larger, more general purpose model. To observe this behaviour, try generating a few tokens after a non-trivial word like "Biology". The model typically treats words that did not frequently appear in training as character names in a story.

All training data is English. As such, input with other languages is out of distribution, and will result in the model treating previous input as character names, ignoring it entirely, or generating meaningless tokens.

Training and evaluation data

Trained for 3 epochs on the TinyStories V2 dataset, produced by GPT-4.

Training procedure

Trained for 400k steps (~7 hours) on 2xH100 80GB PCIe with 32vCPU and 500GB RAM on Runpod.

To replicate, download GPT-4 V2 version of the TinyStories dataset alongside HuggingFace's train_clm.py script. Then run the following:

#! /bin/bash

python train_clm.py \
    --model_type=gpt2 \
    --config_overrides=n_embd=64,n_layer=8,n_head=16 \
    --tokenizer_name=gpt2 \
    --train_file="data/TinyStoriesV2-GPT4-train.txt" \
    --validation_file="data/TinyStoriesV2-GPT4-valid.txt" \
    --block_size=256 \
    --preprocessing_num_workers=8 \
    --output_dir="out" \
    --logging_dir="./log" \
    --logging_steps=100 \
    --logging_strategy=steps \
    --save_steps=5000 \
    --save_total_limit=10 \
    --do_train

Training hyperparameters

The following hyperparameters were used during training:

n_embd: 64
n_layer: 8
n_head: 16
learning_rate: 5e-05
train_batch_size: 16
eval_batch_size: 16
seed: 42
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: linear
num_epochs: 3.0

Framework versions

Transformers 4.35.0.dev0
Pytorch 2.0.1+cu118
Datasets 2.14.5
Tokenizers 0.14.1

calum
/

tinystories-gpt2-3M