bloom-chatml-id / README.md
haidlir's picture
Update README.md (#2)
61ea9c1 verified
metadata
datasets:
  - FreedomIntelligence/alpaca-gpt4-indonesian
  - FreedomIntelligence/evol-instruct-indonesian
  - FreedomIntelligence/sharegpt-indonesian
  - jakartaresearch/indoqa
language:
  - id
library_name: transformers
pipeline_tag: text-generation

Notebook Info

Reference:

Task: Chat or Conversational

Input: User's prompt containing chat templated text in string format

Output: Model's generated text in string format

Experiment:

  • Use bos_token and eos_token to replace <|im_start|> and <|im_end|> in ChatML. (Inspired by: https://asmirnov.xyz/doppelganger)
  • Use left padding and left truncation to conform to max_length.
  • Set max_length = 256 in the training process, which consumes 33.7 GB of memory.

Notebook: