Training Client CLI

nupic_train.main()

Parse commandline and excute fine-tuning example.

nupic_train.run_finetuning(url: str, train_path: str, test_path: str, task_type: str, epochs: int, learning_rate: float, model: str, batch_size: int, seed: int = -1, cache_dir: str = None, overwrite_cache: bool = False, wandb_api_key: str = '', wandb_project_name: str = '', wandb_upload_files: bool = False, hf_access_token: str = '', use_bf16: bool = False, use_lora: bool = False, lora_r: int = 16, lora_alpha: int = 32, lora_dropout: float = 0.05, download_dataset: bool = False, dataset_name: str = '', input_prompt: str = '', response_prompt: str = '', max_steps: int = 100)

Provide CLI to run finetuning job using NuPIC Training Server.

Example:

python -m nupic.client.nupic_train \
    --url [http://localhost:8321](http://localhost:8321) \
    --train_path <path_to_train_set> \        
    --test_path <path_to_test_set> \
    --task_type classification \
    --epochs 5 \
    --learning_rate 1e-5 \
    --model “nupic-sbert.base-v3” \
    --batch_size 32

• Parameters:
  • url (str) – URL of the API server.
  • train_path (str) – path to the train data file.
  • test_path (str) – path to the train data file.
  • task_type (str) – task type. Options are: classification, qa.
  • epochs (int) – number of epochs.
  • learning_rate (float) – learning rate.
  • model (str) – model name.
  • batch_size (int) – batch size.
  • seed (int , optional) – random seed, defaults to -1
  • cache_dir (str , optional) – cache directory for QA, None for no cache, defaults to None
  • overwrite_cache (bool , optional) – overwrite existing cache for QA, defaults to False
  • wandb_api_key (str , optional) – WandB API key, defaults to “”. Required for WandB
  • wandb_project_name (str , optional) – WandB project name, defaults to “”. Required for WandB
  • wandb_upload_files (bool , optional) – To upload model and results to WandB, defaults to False
  • hf_access_token (str , optional) – Hugging Face access token for models that require login.
  • use_bf16 (bool , optional) – whether the LLM should be loaded using bf16 for train.
  • use_lora (bool , optional) – whether the LLM should be using peft with Lora.
  • lora_r (int , optional) – Lora attention dimension (the “rank”).
  • lora_alpha (int , optional) – The alpha parameter for Lora scaling.
  • lora_dropout (float , optional) – The dropout probability for Lora layers.
  • download_dataset (bool) – dataset will be downloaded instead of uploaded.
  • dataset_name (str) – name of the dataset to be downloaded from hf datasets repo.
  • input_prompt (str) – custom input prompt to be presented before the “text” column.
  • response_prompt (str) – custom response prompt to be presented before the “label”.
  • max_steps (int) – number of train steps for LLMs.