StoryLM-10M
A ~12.6M parameter LLaMA-style causal language model trained from scratch on TinyStories.
Architecture
| Parameter | Value |
|---|---|
| d_model | 256 |
| n_heads | 8 (MHA) |
| n_kv_heads | 8 |
| n_layers | 8 |
| head_dim | 32 |
| FFN dim | 1024 (SwiGLU, 4x) |
| Vocab size | 8192 |
| Context length | 512 |
| Tied embeddings | Yes |
| Total params | 12,603,648 |
Training
- Data: TinyStories (2,119,719 stories, 533.7M tokens)
- Steps: 24,400 (batch 4, seq 512)
- Optimizer: AdamW, LR 3e-4, cosine schedule, 500-step warmup
- Best val loss: 1.8635
- Val perplexity: 6.45
Generation Samples
Prompt: Once upon a time, there was a little
Output: Once upon a time, there was a little
Prompt: The cat sat on the
Output: The cat sat on the
Prompt: In the beginning, the world was
Output: In the beginning, the world was
Prompt: A small robot named
Output: A small ro bot named
Prompt: Every morning, the sun
Output: Every morning, the sun
Notes
- Trained on a single RTX 5090 (32 GB) in ~281 seconds of training time.
- The model is a custom implementation (not HuggingFace transformers-compatible out of the box).
- Weights are stored as a single PyTorch checkpoint (
model.pt).