Skip to main content
Version: Next

Tracking Project State

After you run:

constellaxion init

Constellaxion generates a job.json file that captures the complete state of your deployment or training workflow. This file is automatically referenced when running commands like train, serve, deploy, or prompt.

Heads up!

job.json is Constellaxion's memory of what you’re working on — don’t edit it manually unless you know what you’re doing.

💡 job.json Explained​

Example​

{
    "model": {
        "model_id": "my-sentiment-llm",
        "base_model": "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
    },
    "dataset": {
        "train": {
            "local": "./train.csv",
            "cloud": "my-sentiment-llm/data/train.csv"
        },
        "val": {
            "local": "./val.csv",
            "cloud": "my-sentiment-llm/data/val.csv"
        },
        "test": {
            "local": "./test.csv",
            "cloud": "my-sentiment-llm/data/test.csv"
        }
    },
    "training": {
        "epochs": 1,
        "batch_size": 4,
        "tensorboard_url": "https://<tensorboard-url>"
    },
    "deploy": {
        "provider": "gcp",
        "project_id": "<your-project-id>",
        "region": "<your-region>",
        "bucket_name": "<your-bucket-name>",
        "staging_dir": "my-sentiment-llm/staging",
        "experiments_dir": "my-sentiment-llm/experiments",
        "model_path": "my-sentiment-llm/model",
        "service_account": "constellaxion-admin@<your-project-id>.iam.gserviceaccount.com"
    }
}

🔍 Field-by-Field Breakdown

model​

model_id string - The unique label for this project or experiment.

base_model string - The Hugging Face model ID used as the starting point for deployment or fine-tuning.

dataset​

Holds references to your dataset files, both locally and in cloud storage.

Each entry (train, val, test) has:

  • local string - The path on your local file system (usually in the same directory as model.yaml)
  • cloud string - The upload path inside your GCS bucket

These paths are used during job upload and training orchestration.

training​

This section only appears if you’re fine-tuning a model.

  • epochs int - Number of epochs for training
  • batch_size int - Batch size used in training
  • tensorboard_url string (optional) - A link to the TensorBoard experiment tracking your training run, generated automatically on Vertex AI if applicable.

deploy​

This section configures your cloud deployment.

  • provider string - The cloud backend used. Currently supports gcp. (Multi-cloud support coming soon.)
  • project_id string - Your GCP project identifier.
  • region string - GCP region where jobs run.
  • bucket_name string - Your GCS bucket for model artifacts, datasets, logs.
  • staging_dir string - Temp directory where data is uploaded before processing.
  • experiments_dir string - Where logs, checkpoints, and metrics are saved during training.
  • model_path string - Where the final trained model will be saved in GCS.
  • service_account string - The service account used to launch Vertex AI jobs. Must have sufficient IAM permissions.

🧠 Why job.json Matters​

  • It keeps your commands short — no need to re-specify paths or settings
  • It ensures reproducibility — your job context is recorded
  • It’s updated incrementally — e.g., after train, new fields like tensorboard_url or checkpoints may be added

🛠️ Troubleshooting

  • Missing field? Run constellaxion init again to regenerate.
  • Wrong cloud config? Update your model.yaml and re-init.
  • Multiple environments? Keep separate directories with their own job.json files.

🔗 Related Pages​