tinystories-gpt-51m

tinystories-gpt-51m

A 51.2M parameter GPT-2 style language model pretrained from scratch on TinyStories. It writes simple, mostly coherent children's stories in the style of its training corpus, and essentially nothing else.

This model is the artifact of a learning project (gpt-spark) whose goal was to expose every moving part of a real pretraining run in one readable file: data packing, gradient accumulation, mixed precision, LR scheduling, evaluation and checkpointing. The transformer itself is unmodified Hugging Face GPT2LMHeadModel — deliberately, so that the training loop is the thing being learned rather than reimplemented.

It is published for inspection and teaching, not because it is a useful model.

Model details

Half the parameter count is the token embedding table (50257 × 512 = 25.7M) — at this scale, vocabulary dominates. Each of the 8 transformer blocks is only ~3.15M parameters.

For comparison with the smallest published GPT-2:

Usage

from transformers import pipeline

gen = pipeline("text-generation", model="rprouse/tinystories-gpt-51m")
print(gen("Once upon a time", max_new_tokens=180, do_sample=True,
          temperature=0.8, top_k=50, top_p=0.95)[0]["generated_text"])

Or explicitly:

from transformers import GPT2LMHeadModel, GPT2TokenizerFast

repo = "rprouse/tinystories-gpt-51m"
tok = GPT2TokenizerFast.from_pretrained(repo)
model = GPT2LMHeadModel.from_pretrained(repo).eval()

ids = tok("Once upon a time", return_tensors="pt").input_ids
# Position embeddings cover exactly 512 slots: prompt + new tokens must fit inside.
n = min(180, model.config.n_positions - ids.shape[1])
out = model.generate(ids, max_new_tokens=n, do_sample=True, temperature=0.8,
                     top_k=50, top_p=0.95, pad_token_id=tok.eos_token_id)
print(tok.decode(out[0], skip_special_tokens=True))

The context window is a hard limit. GPT-2 uses learned position embeddings, so there are exactly n_positions = 512 of them. Requesting prompt + max_new_tokens beyond 512 indexes off the end of wpe. On CUDA that raises an error; under ROCm's flash-attention kernel it was observed to abort the process rather than raise a clean Python exception. Clamp max_new_tokens as shown above.

Training

Data

The full TinyStories corpus, tokenised once with the GPT-2 BPE tokenizer and packed nanoGPT-style: every document gets <|endoftext|> appended, everything is concatenated into one flat uint16 array per split, and batches are drawn by slicing 512-token windows at random offsets.

There are no document boundaries within a batch and no padding: every token position is a valid training example, and the model sometimes learns to read across an EOS. That is the deliberate tradeoff — maximum data efficiency for slightly noisy document boundaries.

Hyperparameters

Weight decay is withheld from LayerNorm gains on purpose: decaying a gain pulls it toward zero, which fights the normalisation it exists to perform.

Hardware

Trained on a single AMD Strix Halo EVO-X2 (Ryzen AI Max+ 395, gfx1151) on Windows using AMD's ROCm PyTorch builds. Roughly 12 hours wall clock, in two 15,000-step runs of about 6 hours each.

A note on the second half of training

The run was done in two parts: 15,000 steps, then --resume --steps 30000. Because --steps also defines the cosine schedule's horizon, resuming built a fresh 30,000-step cosine and loaded the saved scheduler state into it at step 15,000. The learning rate therefore jumped from ~0 (the end of the first cosine) back up to 5.13e-4 (the midpoint of the new one) and decayed a second time.

The measured cost: validation loss was 1.319 at step 15,000 and 1.308 at step 30,000. The second 15,000 steps — half the total compute — bought roughly 0.01 nats. The model spent them relearning ground it had already covered. A single 30,000-step run under one continuous cosine would very likely have done better for the same cost. This is documented rather than quietly re-run because it is the more instructive artifact.

Evaluation

A single deterministic pass over the entire packed validation split: non-overlapping 512-token windows, token-weighted mean negative log-likelihood. Reproduce with eval_ppl.py:

uv run eval_ppl.py --out ckpt_8x512

For context, the training loop logs a cheaper estimate — the mean over 20 random batches — which across the final ten checkpoints ranged 1.287 to 1.321. The full-set number lands inside that band, so the sampled estimate was unbiased but far too noisy to quote a single best value from. The full-set figure is the one reported here.

No downstream benchmarks were run. Perplexity on the corpus a model was trained on measures fit to that corpus and nothing more.

Sample outputs

All generated with temperature=0.8, top_k=50, top_p=0.95, seed=1337.

Once upon a time, there was a little girl named Lily. She loved to play outside in the sunshine. One day, she went to the park with her mommy and saw a beautiful butterfly. "Wow, mommy! Look at the pretty butterfly!" she said.

Suddenly, Lily saw a man who was very impatient. He was trying to catch the butterfly to catch it. "Mommy, why is he so impatient?" she asked. "He wants to catch the butterfly," her mommy replied.

It was a cold fall morning. The wind was blowing and it was bitter outside. The little girl was sad and started to cry.

Suddenly, a rainbow appeared in the sky! It was so beautiful and colourful. The little girl stopped crying and smiled. […] She felt so happy that she hugged her mum. Then she skipped around the park one last time, feeling safe and warm inside her mum's arms.

Note the flaws visible even in the good samples: "trying to catch the butterfly to catch it", and the "so happy … so happy" repetition. Short-range fluency is solid; phrase-level redundancy is not.

Given an out-of-domain prompt, the model does not decline or hedge — it converts the prompt into a TinyStories fable:

In 1789, the French Revolution began because the family had to move away from the shore.

One day, a brave little boy decided to take the family on an adventure. He was determined to help them, and they followed him. […] From that day on, the family lived happily ever after on the shore.

This is the single most useful thing to understand about the model, and no perplexity number would have told you it.

Intended use

In scope:

• Teaching and studying the mechanics of a pretraining run — a real, small, fully reproducible checkpoint to load, inspect, probe and fine-tune.
• A baseline or starting point for TinyStories-scale experiments: tokenizer ablations, model-shape sweeps, distillation, and interpretability work where a 51M-parameter model that runs on a laptop is a feature.
• Generating simple English children's stories, in the narrow style of the corpus, for demos and tests.

Out of scope:

• Any factual, informational or advisory use. The model has no knowledge base; the French Revolution sample above is what "answering a question" looks like.
• Instruction following, chat or dialogue. This is a base LM with no instruction-tuning or RLHF of any kind.
• Any production or user-facing deployment, and any use where a plausible-sounding wrong output carries a cost.
• Content aimed at children without human review, despite the corpus. Being trained on children's stories is not the same as being safe for children.

Limitations and biases

• The domain is extremely narrow. TinyStories is synthetic, deliberately simple English pitched at a 3–4 year old vocabulary. The model has never seen code, code switching, technical prose, long-form argument or adult register, and its outputs collapse toward simple narrative regardless of the prompt.
• No world knowledge. It was trained on ~474M tokens of synthetic fiction. Anything factual it emits should be assumed false.
• 512-token context, with the hard cutoff described under Usage.
• Repetition and phrase looping, visible in the samples above.
• The tokenizer is oversized for the corpus. The GPT-2 BPE vocabulary was reused unchanged, so 50.2% of the parameters sit in an embedding table covering CJK, emoji and other scripts that never appear in TinyStories. Training a corpus-specific BPE vocabulary is the most promising next change: it would move a large share of the parameter budget out of the embedding table and into the transformer blocks.
• It inherits the biases of its corpus, and of the model that generated it. TinyStories was synthesised by GPT-3.5 and GPT-4, so this model is a distillation of those models' distribution over children's stories — including whatever skews in names, family structures, gender roles and settings that carries. No bias evaluation was performed.
• No safety training. Nothing prevents the model from producing unsuitable output given an adversarial prompt.

Reproducing

git clone https://github.com/rprouse/gpt-spark && cd gpt-spark
uv sync
uv run gpt_train.py --steps 30000          # one continuous cosine — see the note above
uv run gpt_train.py --sample "Once upon a time"
uv run eval_ppl.py                         # the validation figures reported above
uv run tensorboard --logdir runs

The training loop, its rationale and a full command-line reference are in the project README.

Licence and attribution

• Weights and training code: MIT.
• Training data: TinyStories by Ronen Eldan and Yuanzhi Li, licensed CDLA-Sharing-1.0. Users of this model should review that licence with respect to their own use of the data.
• Tokenizer and architecture: GPT-2, by OpenAI, via Hugging Face transformers (MIT).

@misc{eldan2023tinystories,
  title         = {TinyStories: How Small Can Language Models Be and Still Speak Coherent English?},
  author        = {Ronen Eldan and Yuanzhi Li},
  year          = {2023},
  eprint        = {2305.07759},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL}
}