adamroberts/tinystories-5090

llm.kittens TinyStories 124M BF16

This is a 124M-parameter GPT-2-style causal language model trained from scratch on TinyStories with the llm.kittens C++/CUDA trainer, which is a fork of Kaparthy's llm.c with some optimisations for SM120, and multi-stack kernel optimisations.

The model is published as a standard Hugging Face Transformers checkpoint with BF16 safetensors weights. It was trained on a single RTX 5090 in 12 hours.

This isn't a chat model. It's pretrained only, use the Python example to use it.

Result

• Model weights: model.safetensors
• Training step: 20000 / 20000
• Final train loss: 0.785740
• Final validation loss: 0.875080
• Final throughput: 207135 tokens/s
• Final step time: 2531.04 ms
• Final reported BF16 MFU: 39.7%
• Average iteration time: 2605.014347 ms
• Safetensors size: 248,894,656 bytes
• Parameter count: 124,475,904

The TinyStories paper reports eval losses of 1.33 to 1.58 for the 768-hidden-size 1- and 2-layer attention-head ablations in Figure 24. This run's 0.875080 validation loss is lower, but the comparison is not apples-to-apples: this model is a 12-layer GPT-2-style model using GPT-2 tokenization, a 1024-token context, and a different implementation/training setup.

Architecture

• Family: GPT-2-style decoder-only Transformer
• Descriptor: d12
• Layers: 12
• Attention heads: 12
• Hidden size: 768
• Context length: 1024
• Vocabulary size: 50,257
• Precision: BF16 weights

Training

The run used the TinyStories GPT-2 dataset files generated by dev/data/tinystories.py in llm.kittens.

./train_gpt2cu \
    -i "dev/data/tinystories/TinyStories_train.bin" \
    -j "dev/data/tinystories/TinyStories_val.bin" \
    -o "log124M/5090_S" \
    -v 250 -s 20000 -g 144 \
    -h 0 \
    -b 64 -t 1024 -d 524288 \
    -r 0 \
    -z 1 \
    -c 0.1 \
    -l 0.0006 -q 0.0 -u 700 -n 5000 \
    -y 0 \
    -e "d12" \
    -x 20000

Key settings:

• Hardware target: RTX 5090 / SM120
• Micro batch: 64
• Sequence length: 1024
• Total desired batch size: 524,288 tokens
• Max steps: 20,000
• Optimizer: AdamW as implemented in llm.kittens
• Peak learning rate: 6e-4
• Scheduler: cosine
• Warmup: 700 steps
• Final LR fraction: 0.0
• Weight decay: 0.1
• Recompute: off
• ZeRO stage: 1
• Checkpoint interval: 5000 steps

Sample

Prompt/sample emitted at the final checkpoint:

Once upon a time, there was a little boy named Timmy. Timmy loved going to school and playing with his friends. One day, Timmy woke up and felt very hot. He asked his mom if his head hurt. His mom said it might be burnt. Timmy's mom recommended they switch their shirts outside so he would feel better.

Timmy went outside and saw his friends playing. He wanted to join them, but he remembered his mom's recommendation. He switched his shirt right away and felt much cooler. Timmy was happy he listened to his mom and his friends.

Later, during recess, Timmy's friend asked him to go on the slide.

Files

• model.safetensors: BF16 Transformers weights.
• config.json: GPT-2 model configuration.
• generation_config.json: default generation settings.
• tokenizer.json: GPT-2 tokenizer.
• vocab.json and merges.txt: GPT-2 BPE vocabulary files.

Loading

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "adamroberts/tinystories-5090"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16)

inputs = tokenizer("Once upon a time", return_tensors="pt")
with torch.inference_mode():
    outputs = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.8)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

GGUF usage

This is a GPT-2 completion model, not a chat/instruction model. In LM Studio or llama.cpp, use a plain completion/default preset with no chat template. Always set a finite max token limit; the model was trained for continuation and may not emit <|endoftext|> during normal story generation.

Recommended sampling settings:

• Temperature: 0.8
• Top-p: 0.95
• Top-k: 50
• Repetition penalty: 1.05
• Stop/EOS token: <|endoftext|> / token id 50256

For llama.cpp CLI usage, prefer llama-completion over llama-cli:

llama-completion \
  -m tinystories-5090.Q8_0.gguf \
  -p "Once upon a time" \
  -n 144 --temp 0.8 --top-p 0.95 --top-k 50 --repeat-penalty 1.05

Source implementation: https://github.com/adamdroberts/llm.kittens

TinyStories reference paper: https://arxiv.org/abs/2305.07759