AlexaTM 20B

text
paid
AlexaTM 20B (2022) is Amazon’s 20B-parameter multilingual seq2seq model. It achieves state-of-the-art performance on tas...
Version: 20B
Released: 3y 10m ago on 01/01/2022

Architecture

  • parameters: 20B (encoder-decoder Transformer)
  • context_length: 2048 tokens (estimate)
  • training_data: Proprietary multilingual corpora
  • inference: Seq2seq Transformer

Capabilities

  • Multilingual text generation and understanding
  • Strong performance on summarization, translation, and QA tasks
  • Supports a dozen languages (English, Spanish, Chinese, etc.)

Benchmarks

  • SuperGLUE: Outperforms GPT-3 (175B) on SuperGLUE (zero-shot)
  • XNLI: State-of-the-art on XNLI multilingual benchmark

Safety

  • Proprietary model with internal safety measures
  • details not publicly disclosed.

Deployment

  • regions: global (AWS regions)
  • hosting: AWS SageMaker
  • integrations: Available via SageMaker JumpStart

Tags

multilingualseq2seqNLPAmazon

Join our community

Connect with others, share experiences, and stay in the loop.