0rn0/llama2-15m-tinystories-sft

Llama 2 15M — TinyStories SFT

Model Details

• Architecture: Llama 2 (custom implementation)
• Parameters: ~15.2M
• Context Length: 256 tokens
• Embedding Dim: 288
• Attention Heads: 6
• KV Heads: 6 (Multi-Head Attention)
• Transformer Layers: 6
• Tokenizer: Llama 2 SentencePiece (vocab size 32,000)

Training

Pre-training

Pre-trained from scratch on the TinyStories dataset. Pre-trained weights available at 0rn0/llama2-15m-tinystories.

Supervised Fine-Tuning (SFT)

Fine-tuned for 3 epochs on 0rn0/tinystories-instruct-balanced, a curated instruction-tuning dataset with balanced happy/sad endings (50-50 split, ~325K examples).

SFT Hyperparameters:

• Optimizer: AdamW
• Learning Rate: 5e-5
• Batch Size: 64
• Epochs: 3
• Weight Decay: 0.1
• Max Sequence Length: 256

Training Results:

• Best Validation Loss: 1.1275 (Epoch 3)
• Final Training Loss: 1.0855
• Training Time: ~6 hours (1x A100 GPU)

Usage

This is a custom PyTorch model (not transformers-compatible). Requires source code from the GitHub repository.

git clone https://github.com/aryandeore/llama2-sft.git
cd llama2-sft
uv sync

Instruction Format

Write a story about: {topic}
With: {ending} ending

### Story:

Where {ending} is either happy or sad.

Load Model

from models.llama2 import Transformer

model = Transformer.from_pretrained("0rn0/llama2-15m-tinystories-sft")
model.eval()

Generate Story

uv run python generate.py \
  --topic "a brave knight" \
  --ending sad \
  --temperature 0.7 \
  --max-tokens 192

Examples

Happy ending

Prompt: Write a story about: a little girl and her dog

Once upon a time, there was a little girl named Sophie who had a small dog named Buddy. Every morning, Sophie and Buddy would run to the park together. One sunny day, Sophie accidentally dropped her favorite toy near the big oak tree. Buddy immediately began to sniff around the grass, searching everywhere. Finally, Buddy found it! Sophie was so happy and hugged her dog tightly. From that day on, Sophie knew that Buddy would always be there to help her find anything she lost. Sophie and Buddy played together in the park every single day, and they were the very best of friends forever.

Sad ending

Prompt: Write a story about: a boy who lost his favorite toy

Once there was a young boy named Thomas who loved his wooden toy train. He played with it every single day and took it everywhere he went. One afternoon, Thomas left his beloved toy train at the park by accident. When he realized it was missing, he rushed back to search for it. He looked everywhere but could not find his precious toy. Thomas felt very sad and disappointed. His parents tried to comfort him, but nothing seemed to help. Thomas went to bed that night with tears in his eyes, wishing he could have his toy train back.

Limitations

• Generates short stories only (~150-200 words)
• Trained on synthetic data from TinyStories
• Limited vocabulary and narrative complexity
• Binary ending classification (happy/sad only)
• Small context window (256 tokens)

Links

• Pre-trained model: 0rn0/llama2-15m-tinystories
• Dataset: 0rn0/tinystories-instruct-balanced
• GitHub: AryanDeore/llama2-sft
• Collection: Tiny Stories
• Related GPT-2 Models: 0rn0/gpt2-30m-tinystories-sft, 0rn0/gpt2-125m-tinystories-sft