tinystories-15m — a language model pretrained entirely by volunteers

Training complete (August 2, 2026). Final state: outer step 195, val loss 2.872, ~423M community tokens — pinned at revision 5c39e2cb. This model stays fully usable (snippet below). The co-op's current run is fineweb-150m — a 145M-param model on FineWeb-Edu; npx coop-ai start now contributes there.

This 14.8M-parameter model was pretrained from scratch with no cluster, no server, and no funding: volunteers ran training rounds on their own computers and submitted compressed pseudo-gradients as pull requests on a public Hugging Face dataset repo. A stateless GitHub Actions cron job aggregated them into outer steps — DiLoCo-style low-communication data parallelism, coordinated by nothing but free-tier infrastructure.

• Training run: July 27 – August 2, 2026 · 195 outer steps · ~423M tokens (≈121% of the ~300M Chinchilla-optimal budget for this size)
• Final validation loss: 2.872 cross-entropy on held-out TinyStories (≈17 perplexity over an 8k vocab; random-init baseline is ln 8192 ≈ 9.01)
• Contributors: 4 identities on consumer hardware — Apple Silicon laptops, plain CPUs, and one GPU (public leaderboard)
• Everything is auditable: every gradient arrived as a public PR on the inbox repo, every outer step is a revision of this repo, and the credit ledger lives on the ledger branch

Sample output

Prompt: Once upon a time

Once upon a time, there was a little girl named Lily. She loved to play outside in her garden. One day, she found a big, yellow flower. She was so happy and started to look at the pretty petals. She wanted to decorate her garden with the flowers and show them to her mom.

But when Lily went to visit her grandma, she saw her small flower. She looked guilty and felt sad. Lily wanted to show her flower to her grandma, but her flower was gone. She started to cry because she missed her flower.

Model details

How it was trained

Workers download the current checkpoint, run up to 500 local AdamW steps (lr 3e-4, betas 0.9/0.95, weight decay 0.1, grad clip 1.0) on their personal TinyStories shard, and submit the pseudo-gradient θ_outer − θ_local, int8-quantized, as a pull request. Each aggregation tick is stateless: it drops over-stale submissions (> 8 steps old), L2-clips each delta, cosine-gates against the clipped weighted mean, merges same-user submissions into one token-weighted vote, robust-aggregates the votes (20% trimmed mean), and takes one Nesterov outer step (lr 0.7, momentum 0.9). Contributions are Byzantine-filtered but the compute itself is unverified — this is a volunteer-trust experiment as much as a model.

How to use

The model uses its own tiny GPT class (not transformers):

# pip install git+https://github.com/commonsense-ai/coop
import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from tokenizers import Tokenizer
from coop.model import GPT, GPTConfig, load_canonical_state

repo = "commonsense-ai/tinystories-15m"
model = GPT.from_config(GPTConfig())  # defaults match this checkpoint
load_canonical_state(model, load_file(hf_hub_download(repo, "checkpoint.safetensors")))
model.eval()

tok = Tokenizer.from_file(hf_hub_download(repo, "tokenizer.json"))
idx = torch.tensor([tok.encode("Once upon a time").ids])
out = model.generate(idx, max_new_tokens=200, temperature=0.8, top_k=50)
print(tok.decode(out[0].tolist()))

Limitations

It writes toddler fiction, and only toddler fiction. TinyStories is a synthetic corpus with a ~1,500-word vocabulary world, so the model has no knowledge, no instruction-following, and no register other than bedtime stories about children, animals, and feelings. Plots meander and pronouns drift. 512-token context. It is an educational artifact demonstrating that strangers on the internet can pretrain a real model together — use it to study that, not to ship products.

Contribute

The run may continue and successors are planned. Joining takes one command:

npx coop-ai start

Code, architecture writeup, and the leaderboard: github.com/commonsense-ai/coop