Text Gen Fine-tuning CLI Reference

A reference guide for text generation fine-tuning CLI commands and their behaviors. This document provides a detailed description of every command of the OctoAI Text Generation Fine-tuning CLI. Remember that you can install the CLI by following CLI & SDK Installation.

octoai tune text

Create text generation fine-tuning jobs.

Create text generation fine-tuning jobs by dataset name:

1
octoai tune text --checkpoint <CHECKPOINT_NAME> --dataset <DATASET_NAME> --name <TUNE_NAME>

Create text generation fine-tuning jobs by dataset ID:

1
octoai tune text --checkpoint <CHECKPOINT_NAME> --dataset <DATASET_ID> --name <TUNE_NAME>

Flags

• -b, --checkpoint <string> (required) This helps the tuning job to relate to the model. Checkpoint for Llama 3.1 8B Instruct is “octoai:meta-llama-3-1-8b-instruct”. Checkpoint for Llama 3.1 70B Instruct is “octoai:meta-llama-3-1-70b-instruct”. For Llama 3.1 70B Instruct access, please contact us.
• -f, --dataset <strings> (required) Datasets used for text fine-tuning. Datasets can be retrieved by using asset IDs or names. It is possible to submit multiple datasets at one time.
• -n, --name <string> (required) Name of the text generation fine-tuning. It should be unique.
• -l, --learning-rate <float32> (optional) The learning rate. Default is 0.0001.
• -r, --lora-rank <int32> (optional) Rank of LoRa (Low Rank Adaptation). Default is 8.
• -e, --epochs <int32> (optional) Number of epochs. Default is 3.
• --wandb-project-name <string> (optional) Weights & Biases project name.
• --wandb-secret-key <string> (optional) Name of OctoAI secret containing Weights & Biases token. Note: This is the “key” of OctoAI secret, not the value. Set up your Weights & Biases account by following Weights & Biases Setup Guide.
• -d, --description <float32> (optional) Description of text generation fine-tuning.

octoai tune list

List all of the fine-tuning jobs. It currently lists 4 columns: tune_id, tune_name, tune_type, and status of tuning.

1
octoai tune list

List all of the text generation fine-tuning jobs.

1
octoai tune list -t text_tune

Flags:

• -t, --type <string> List the fine-tuning jobs of a specific type.
• -n, --name <string> List the fine-tuning with a specific tune_name.
• -b, --checkpoint <string> This helps the tuning job to relate to the model. Checkpoint for Llama 3.1 8B Instruct is “octoai:meta-llama-3-1-8b-Instruct”. Checkpoint for Llama 3.1 70B Instruct is “octoai:meta-llama-3-1-70b-Instruct”.

octoai tune get

Lists the detailed information a specific fine-tuning job in 4 columns: tune_id, tune_name, tune_type, and status.

• Get job status by ID

1
 octoai tune get —-id <DATASET_ID>

• Get job status by name

1
 octoai tune get —-name <DATASET_NAME>

• Get job detail by ID in json format

1
 octoai tune get —-id <DATASET_ID> -o json

Flags

• -i, --id <string> Retrieve the fine-tuning information for specific tune_id. You can locate the tune_id in the first column of the output table generated by the command octoai tune list.
• -n, --name <string> Retrieve the fine-tuning information for a specific tune_name. You can locate the tune_name in the first column of the output table generated by the command octoai tune list.
• -o, --output-format <string> Specify the output format for the fine-tuning information listing. There are two options: table(by default) and json. 1) Table: It lists information in table format. 2)Json: Select this option to receive the information in JSON format, which includes additional details beyond the basic 4 columns.

octoai tune delete

This command is used to delete a fine-tuning job.

• -i, --id <string> Delete the fine-tuning with a specific tune_id.

1
 tune delete --id <TUNE_ID>

• -n, --name <string> Delete the fine-tuning with a specific tune_name.

1
 tune delete --name <TUNE_NAME>

octoai tune cancel

This command is used to cancel an ongoing fine-tuning job.

• -i, --id <string> Cancel the fine-tuning with a specific tune_id.

1
  tune cancel --id <TUNE_ID>

• -n, --name <string> Cancel the fine-tuning with a specific tune_name.

1
  tune cancel --name <TUNE_NAME>