GPT-J

• 📙Paper: GPT-J
• 📚Publisher: other
• 🏠Author Affiliation: EleutherAI
• 🔑Public: ✅
• 🌐Architecture
  • Encoder-Decoder
  • Decoder-Only
• 📏Model Size
  • 6B
• 🗂️Data pre-processing
  • Data Resource
    • The Pile
  • De-duplication: /
  • Filter Strategies
    • /
• 🍉Tokenizer
  • Technology
    • Byte-level Byte-Pair-Encoding (BBPE)
    • SentencePiece
  • Details
    • Same tokenizer as GPT-2/3 (origin words)
• 🧪Hyperparameters (GPT-J 6B)
  • optimizer: Adam
    • betas: /
    • eps: /
  • batch size: /
  • context window: 2,048
  • gradient accumulation steps: 16
  • warmup steps: 3,000
  • learning rate: /
  • weight decay: 0.1
  • decay schedule
    • Cosine
    • Linear
    • Polynomial
    • Inverse Square
  • precision floating point: bf16
• 🏃‍♀️Training
  • model initialization: /
  • training strategies
    • left-to-right
    • fill-in-the-middle
  • trained tokens/steps: /
  • hardware: TPU Research Cloud
  • training time: /