trixyL's picture
update: metadata card
5e22d07 verified
|
Raw
History Blame Contribute Delete
1.09 kB
metadata
license: apache-2.0
datasets:
  - trixyL/simplestories-8k-megatron
language:
  - en
pipeline_tag: text-generation
tags:
  - simplestories
  - '512'
  - 8k
  - transformerlm

🧠✨ TransformerLM (AR 512, 8K vocab) — SimpleStories

This is the result of the code from https://github.com/triloy8/transformerlm, a minimal autoregressive Transformer LM trained on SimpleStories with a 512-token context and an 8K vocab tokenizer. ✨

✅ Key Facts

  • Model type: Autoregressive Transformer LM
  • Dataset: SimpleStories
  • Context length: 512 tokens
  • Tokenizer vocab size: 8,000
  • Layers: 12
  • Heads: 8
  • d_model: 512
  • d_ff: 2,048
  • Training setup: Single NVIDIA A40 48GB
  • Runtime: ~20 hours ⏱️

📦 What’s Inside

  • 60k steps from a 60k run, including:
    • Optimizer state
    • RNG state
    • Safetensors weights
  • Tokenizer config
  • Run config

🚀 Reproducibility

To reproduce the run:

Exact commit that launched the train: https://github.com/triloy8/transformerlm/commit/06cb4831d47c04a18573bee8e28dc83b10086d06