CodeRL

• 📙Paper: CodeRL: Mastering Code Generation through Pretrained Models and Deep Reinforcement Learning
• 📚Publisher: NeurIPS
• 🏠Author Affiliation: Salesforce Research
• 🔑Public: ✅
• 🌐Architecture
  • Encoder-Decoder
  • Decoder-Only
• 📏Model Size
  • 770M
• 🗂️Data pre-processing
  • Data Resource
    • APPS
  • De-duplication: /
  • Filter Strategies
    • /
• 🍉Tokenizer
  • Technology
    • Byte-level Byte-Pair-Encoding (BBPE)
    • SentencePiece
  • Details
    • We adopt the code-specific tokenizer from Wang et al. [2021].
• 🧪Hyperparameters (CodeRL 770M)
  • optimizer: AdamW
    • betas: /
    • eps: /
  • batch size: 64
  • context window: /
  • gradient accumulation steps: /
  • warmup steps: /
  • learning rate: /
  • weight decay: /
  • decay schedule
    • Cosine
    • Linear
    • Polynomial
    • Inverse Square
  • precision floating point: /
• 🏃‍♀️Training
  • model initialization: CodeT5-large-ntp-py
  • training strategies
    • left-to-right
    • fill-in-the-middle
    • reinforcement learning
  • trained tokens/steps: /
  • hardware: 1 A100 GPU
  • training time: fine-tuned CodeT5-large 30 hours