BLOOM

• 📙Paper: BLOOM A 176B-Parameter Open-Access Multilingual Language Model
• 📚Publisher: arxiv
• 🏠Author Affiliation: BigScience
• 🔑Public: ✅
• 🌐Architecture
  • Encoder-Decoder
  • Decoder-Only
• 📏Model Size
  • 176B
• 🗂️Data pre-processing
  • Data Resource
    • Our pre-training dataset is based on a snapshot of selected public GitHub repositories taken on 2021/07/14.
  • De-duplication: ✅
  • Filter Strategies
    • /
• 🍉Tokenizer
  • Technology
    • Byte-level Byte-Pair-Encoding (BBPE)
    • SentencePiece
  • Details
    • The tokenizer is a learned subword tokenizer trained using BBPE.
• 🧪Hyperparameters (BLOOM 176B)
  • optimizer: Adam
    • betas: 0.9, 0.95
    • eps: 1e-8
  • batch size: 2,048
  • context window: 2,048
  • gradient accumulation steps: /
  • warmup steps: /
  • learning rate: 6e-5
  • weight decay: 0.1
  • decay schedule
    • Cosine
    • Linear
    • Polynomial
    • Inverse Square
  • precision floating point: bp16
• 🏃‍♀️Training
  • model initialization: from scratch
  • training strategies
    • left-to-right
    • fill-in-the-middle
  • trained tokens/steps: 366B tokens
  • hardware: /
  • training time: /