\n<\/p>\n
At this time we’re saying the Google Colab Command-Line Interface<\/a> (CLI), which bridges the hole between your native terminal and distant Colab runtimes, offering a zero-friction execution platform for each builders and AI brokers. The Colab CLI provides:<\/p>\n As a result of the Colab CLI integrates seamlessly into normal terminal environments, it may be utilized by any agent with terminal entry. To make sure your AI assistants can hit the bottom operating, the CLI features a prepackaged Colab ability file<\/a> that gives brokers with prompt, built-in context on precisely how one can leverage the CLI. Let’s take a look at an actual life instance of one thing a person or agent may strive with the Colab CLI.<\/p>\n *Word that whereas the instance under highlights Antigravity<\/a> agent utilizing Colab CLI as a device, Colab CLI can simply be utilized by Claude Code, Codex, and different brokers.<\/p>\n Right here is how an Agent can use the Colab CLI for a real-world ML workflow:<\/p>\n The CLI can be utilized to run an actual QLoRA pipeline that runs end-to-end with only a handful of instructions. Offload heavy computational lifting to a GPU with out typing a single cloud provisioning command by Instructing Antigravity (or your agent of selection) to construct a distant fine-tuning job. On this situation, we ask our agent to make use of the Colab CLI to fine-tune google\/gemma-3-1b-it<\/a> on a Textual content-to-SQL dataset<\/a> to make the mannequin higher at writing SQL queries.<\/p>\n The Antigravity immediate:<\/b> Antigravity executes:<\/b><\/p>\n<\/div>\n \n Shell\n <\/p>\n<\/div>\n Antigravity additionally makes use of the “colab obtain” command to obtain the adapter mannequin, adapter config, tokenizer config, and tokenizer, which can be utilized to load and run your fine-tuned mannequin domestically. Now you could have a remotely fine-tuned mannequin able to serve out of your native gadget!<\/p>\n\n
colab --gpu A100<\/code> or colab --gpu T4<\/code>).<\/li>\ncolab exec<\/code>.<\/li>\n.ipynb<\/code> logs by way of colab obtain<\/code> and colab log<\/code>.<\/li>\ncolab repl<\/code> or colab console<\/code>.<\/li>\n<\/ul>\n<\/div>\nAgent-driven workflows in motion<\/b><\/h3>\n
Nice-tuning Gemma 3-1B<\/b><\/h4>\n
Use the Colab CLI (https:\/\/github.com\/googlecolab\/google-colab-cli) to fine-tune Gemma 3 1B utilizing QLoRA. Provision a Colab T4 GPU occasion, set up the mandatory ML packages (transformers, datasets, peft, trl, and so on.), run my native ~finetune_run<\/a>.py<\/a> script remotely, obtain the ensuing safetensors adapter, save the pocket book log, and cleanup.<\/p>\n$ colab new --gpu T4
\n$ colab set up transformers datasets peft trl bitsandbytes speed up
\n$ colab exec -f finetune_run.py
\n$ colab log --output gemma_finetune_log.ipynb
\n$ colab cease<\/code><\/pre>\nAttempt it out now<\/b><\/h3>\n