TeenyTinyStories-33M

GPT-Neo storyteller (nominal 33M; actual 68,514,048 parameters including the 50,257-token embedding) trained on TeenyTinyStories (ableian/TeenyTinyStories)

Training

• Architecture: GPT-Neo, 4 layers, hidden 768, 16 heads, window 256, context 512, fp32.
• Budget: 20 epochs over written train tokens (15,212 steps x 655,360 tokens = 9.97B tokens), batch 1280 x 512, constant LR 5e-4, weight decay 0.1, betas (0.9, 0.95), clip 1.0.
• Tokenizer: EleutherAI/gpt-neo-125M (with EOS 50256; one EOS per story).

Held-out loss (nats, full files)

• TTS-val (20,000 stories): 1.293 / ppl 3.64
• TS-valid capped to stories of at most 400 tokens (20,711 stories): 1.539 / ppl 4.66
• TS-valid full (21,990 stories): 1.529 / ppl 4.61

Intended use

Research on small-scale synthetic pretraining data. Generates simple children's stories; sample at temperature 0.8-1.0. See the TeenyTinyStories paper for the corpus recipe and judge evaluations.

License

Apache-2.0.