Paths and environment variables inside the container

The GEO_* environment variables and /geo/… paths are exactly the names the platform actually provides.

Directories#

manifest.json#

It sits directly under GEO_DATA_DIR. path is relative to GEO_DATA_DIR, so join the two as is to open a file.

{
  "dataset": { "id": "ds-…", "name": "…" },
  "files": [
    { "path": "images/site_008.jpg", "kind": "image",
      "meta": { "width": 640, "height": 480, "format": "JPEG",
                "label_files": [{ "filename": "site_008.json", "format": "labelme" }] } },
    { "path": "annotations/site_008.json", "kind": "annotation",
      "meta": { "format": "labelme", "annotation_count": 2, "match_rate": 1.0 } },
    { "path": "tabular/features.csv", "kind": "tabular",
      "meta": { "row_count": 300, "columns": [{ "name": "label", "dtype": "object" }] } }
  ]
}

• kind is one of image · annotation · timeseries · tabular · pointcloud, and matches the subdirectory the file is placed in. Only files that passed validation are placed.
• The directory also holds a hidden file the platform uses for its cache check (.stamp). Read the file list from manifest.json instead of scanning the directory.
• The folder structure inside a zip is flattened on upload. If names collide, -2 is appended to the later one. Do not rely on structure such as per-class subfolders.
• meta holds values filled in by the platform validators (image size and paired label files, annotation format, row count and columns of a table, and so on). The example is shortened and may omit keys.
• Labels arrive in a different shape for each kind. Image data gets annotation files (COCO JSON · LabelMe JSON · VOC XML), tabular gets a column inside the CSV, and point clouds get an integer property inside the PLY file.
• timeseries and tabular use the same .csv/.parquet files. The modality chosen when the dataset was created tells them apart. The meta of tabular has no time_column key at all.
• The trainer does the train/val split. The platform does not split the files; it only writes the ratio and seed into split in experiment.json.

Environment variables#

• GEO_HP_* naming rule: hyperparameter keys must match ^[a-z][a-z0-9_]*$, and the environment variable name is that key converted to upper case as is (lr → GEO_HP_LR). Nested structures and lists cannot be expressed as environment variables, so they appear only in experiment.json.
• All values are strings. Convert them to the type of the default value when you read them (see the hyperparameter() function in the example). Booleans arrive as lower-case true/false. In Python bool("false") is true, so compare the string directly.
• Only keys present in the step 7 YAML of the wizard arrive. At submission the keys are checked against the allow list of the chosen framework, so a key the catalog does not have cannot be sent. If the YAML is submitted empty, no GEO_HP_* arrives at all, so keep a default in the image for every value.
• What is not passed in: object storage addresses and keys. The platform fetches the data in advance, so no storage credentials reach the training container.
• The platform does not set PYTHONUNBUFFERED. To see logs in real time, add ENV PYTHONUNBUFFERED=1 to your Dockerfile.

Resource limits#

experiment.json (advanced)#

All the main values also come as environment variables, so most images do not read this file. Read it only when you need the classes list, the split ratio or nested hyperparameters. The file contains no credentials, so it is safe to write it to the log.

{
  "experiment_id": "7abfb12e3a8d…",
  "name": "mnist-external-smoke-3",
  "tenant": "DEMO",
  "created_at": "2026-09-21T06:59:47Z",

  "task": "object_detection",
  "framework": "yolo",
  "model": "yolo26n",
  "classes": ["head", "helmet"],

  "split": { "method": "random", "train_percent": 80, "seed": 0 },
  "hyperparameters": { "epochs": 2 },

  "device": "cpu",
  "gpu": 0,

  "evaluation": { "benchmark": true, "speed_test": false },
  "export": { "onnx": false, "tensorrt": false },
  "serving": { "runtime": "cpu" },
  "register": { "enabled": false, "model_name": "" },
  "pretrained": { "source": "catalog" },

  "dataset": { "id": "ds-…", "name": "…", "file_count": 16,
               "path": "/geo/dataset", "manifest": "/geo/dataset/manifest.json",
               "pointcloud_labels": null },
  "paths": { "data_dir": "/geo/dataset", "work_dir": "/geo/work" },
  "mlflow": { "experiment_name": "mnist-external-smoke-3", "run_id": "…" }
}