Cortex Training Client

Python SDK, command-line tools, runnable recipes, and documentation for the Cortex Training REST API.

Install

Requires Python 3.10 or later and uv. Create an environment and install the client:

uv venv
source .venv/bin/activate
uv pip install git+https://github.com/snowflakedb/cortex-training.git

The package includes the CLI, log TUI, and Python SDK. pip install also works in place of uv pip install in an active Python environment.

Log In

Create ~/cortex-training-config.json using the connection template. Set host to your Snowflake account hostname, pat to your programmatic access token, and database and schema to the location of your Cortex Training endpoint. Keep this file outside the repository and do not commit credentials.

Login validates the config and remembers its path for future CLI commands:

cortex-training login ~/cortex-training-config.json

The equivalent cortex-training login --config ~/cortex-training-config.json form is also supported.

ct is an alias for cortex-training: every CLI example also works with ct. See connection setup for more detail.

Try the CLI

cortex-training capacity           # Check available GPU capacity
cortex-training list               # List jobs
cortex-training get JOB_ID         # Inspect a job from the list
cortex-training tui                # Pick a job and view its logs

See the CLI quick reference for submission, training, generation, checkpoints, and log downloads.

Run a Recipe

Start with Run a Quick SFT Job, a short end-to-end supervised fine-tuning walkthrough.

Recipes require a repository checkout; they are not included in the installed package:

git clone https://github.com/snowflakedb/cortex-training.git
cd cortex-training

Install the recipe dependencies, then run the chosen recipe's commands from the repository root. Recipes take an explicit config=/path/to/config.json argument; use the same connection file you logged in with.

Task	Recipe
Fine-tune a chat model with LoRA or full-parameter training	Conversational SFT
Train math reasoning with reinforcement learning	Math GRPO
Serve a model or checkpoint, generate responses, and evaluate	Inference endpoint

Check cortex-training capacity and the recipe's GPU requirements before starting a run. Browse the recipe index for all workflows.

More Documentation

• Getting started and prerequisites
• CLI commands and configuration
• Python SDK
• REST API
• Model and training-method compatibility
• Job management and cluster status
• RL framework integrations

Development

Editable Install

From a repository checkout, with your Python environment active:

uv pip install -e ".[dev]"

Build a Wheel

With uv installed, run:

./scripts/build_wheel.sh

The script builds the package using pyproject.toml in an isolated build environment and writes the wheel to dist/. It can be invoked from any working directory and does not install the package's runtime dependencies.

Repository Map

Path	Purpose
model-catalog/	Supported models, context limits, and recommended job profiles
docs/	Getting started material, concepts, guides, and reference
recipes/	End-to-end training, sampling, and evaluation workflows
examples/api/	Small JSON examples for individual API operations
examples/config/	Connection configuration templates
src/cortex_training/	Installable Python client
tests/	Client and CLI tests
cluster-status.py	Optional watch view of running jobs and GPU usage