CodeModernBERT-Owl-v1🦉

Model Details

• Model type: Bi-encoder architecture based on ModernBERT
• Architecture:
  • Hidden size: 768
  • Layers: 22
  • Attention heads: 12
  • Intermediate size: 1,152
  • Max position embeddings: 8,192
  • Local attention window size: 128
  • RoPE positional encoding: θ = 160,000
  • Local RoPE positional encoding: θ = 10,000
• Sequence length: up to 2,048 tokens for code and docstring inputs during pretraining
• Implementation: Back-end in Python; integrated into OwlSpotLight, a Visual Studio Code extension.

Pretraining

• Tokenizer: Custom BPE tokenizer trained for code and docstring pairs.
• Data: Functions and natural language descriptions extracted from GitHub repositories.
• Masking strategy: Two-phase pretraining.
  • Phase 1: Random Masked Language Modeling (MLM)
    30% of tokens in code functions are randomly masked and predicted using standard MLM.
  • Phase 2: Line-level Span Masking
    Inspired by SpanBERT, continued pretraining on the same data with span masking at line granularity:
    1. Convert input tokens back to strings.
    2. Detect newline tokens with regex and segment inputs by line.
    3. Exclude whitespace-only tokens from masking.
    4. Apply padding to align sequence lengths.
    5. Randomly mask 30% of tokens in each line segment and predict them.
• Pretraining hyperparameters:
  • Batch size: 20
  • Gradient accumulation steps: 6
  • Effective batch size: 120
  • Optimizer: AdamW
  • Learning rate: 5e-5
  • Scheduler: Cosine
  • Epochs: 2
  • Precision: Mixed precision (fp16) using transformers