CodeGPT

• 📙Paper: CodeXGLUE A Machine Learning Benchmark Dataset for Code Understanding and Generation
• 📚Publisher: NeurIPS
• 🏠Author Affiliation: Microsoft
• 🔑Public: ✅
• 🌐Architecture
  • Encoder-Decoder
  • Decoder-Only
• 📏Model Size
  • 124M
• 🗂️Data pre-processing
  • Data Resource
    • We pretrain monolingual models on Python and Java corpora from the CodeSearchNet dataset.
  • De-duplication: ❌
  • Filter Strategies
    • /
• 🍉Tokenizer
  • Technology
    • Byte-level Byte-Pair-Encoding (BBPE)
    • SentencePiece
  • Details
    • Optional 1: Newly obtained on the code corpus
    • Optional 2: GPT-2 tokenizer
• 🧪Hyperparameters (CodeGPT 124M)
  • optimizer: Adam
    • betas: /
    • eps: /
  • batch size: 32
  • context window: 768
  • gradient accumulation steps: /
  • warmup steps: /
  • learning rate: 5e-5
  • weight decay: /
  • decay schedule
    • Cosine
    • Linear
    • Polynomial
    • Inverse Square
  • precision floating point: /
• 🏃‍♀️Training
  • model initialization: GPT-2
  • training strategies
    • left-to-right
    • fill-in-the-middle
  • trained tokens/steps: /
  • hardware: 2 P100 GPUs
  • training time: /