Skip to main content
Glama
amanharshx
by amanharshx

training_start

Destructive

Start a cloud training job from an existing model or YOLO base checkpoint. Validates the dataset immediately and requires cost confirmation before launching a billable run.

Instructions

Start a cloud training job from an existing model or official YOLO base checkpoint (state-changing, may cost credits). The dataset is validated immediately, so an unusable dataset is rejected before any compute starts. In checkpoint mode a project model is created first, then the checkpoint's task is checked against the dataset's task: with a list of datasets this is checked for every entry before that model is created, but with a single dataset the check happens server-side when training starts, by which point the model already exists — a mismatch there still leaves that model behind, unrequested, and the error names it so it can be reviewed and deleted (models_delete) if unwanted. Starting is billable immediately: the platform has no cost preview before that, so the projected cost and remaining balance are only reported after the job starts. Training an existing model that already has a recorded run (any status past pending/untrained) replaces that model's status, epoch count, and per-epoch metric history the instant the new job starts, and that history cannot be recovered afterward; the previously uploaded weights survive. That path requires confirm_history_loss=true in addition to confirm_cost=true. Checkpoint mode never destroys an existing model's history since it always creates a new one, so it never needs confirm_history_loss — its own risk is the possible leftover model described above. An untrained or never-trained model needs no extra confirmation either. Use training_cancel to stop a job that is already running. Requires confirm_cost=true.

Input Schema

TableJSON Schema
NameRequiredDescriptionDefault
nameNoRun name for callbacks.
batchNoImages per batch: -1 targets about 60% GPU memory, a value between 0 and 1 sets a memory fraction, and a positive integer fixes the image count.
imgszNoTarget input size: square batches normally, or the long-side size with rect=true.
modelYesExisting model ref, or official YOLO base checkpoint such as yolo11n.pt or yolo11n-seg.pt. Checkpoint mode auto-creates a project model.
epochsNoMaximum full passes over the training set.
datasetYesDataset ref by slug, owner/slug, or a ul://owner/datasets/slug URI, or a list of refs to fine-tune on sequentially.
projectYesProject ref that owns the training job and resolved model.
gpu_typeYesCloud GPU type to allocate for training.
train_argsNoAdditional YOLO training arguments passed through to the platform. epochs, imgsz, batch, and name here are silently overridden by the matching top-level input when both are set; data and model are rejected outright if present here.
confirm_costNoMust be true to allow a credit-costing training run. Starting is billable immediately; the platform has no cost preview before that, so the estimated cost and remaining balance are only reported after the job starts.
confirm_history_lossNoMust be true to restart training on an existing model that already has a recorded run. Doing so replaces that model's status, epoch count, and per-epoch metric history irrecoverably; the previously uploaded weights survive. Not required for an untrained model or for checkpoint mode, which creates a new model instead. Separate from confirm_cost.

Schema Changelog

Changes observed during successful MCP inspections.

  1. Changed6 schema fields changedv0.1.13
    • addedInput schema / properties / batch / description
      Added value: +"Images per batch: -1 targets about 60% GPU memory, a value between 0 and 1 sets a memory fraction, and a positive integer fixes the image count."
    • changedInput schema / properties / dataset / description
      Previous value: -"Dataset ref by slug, owner/slug, or ul:// URI, or a list of refs to fine-tune on sequentially."New value: +"Dataset ref by slug, owner/slug, or a ul://owner/datasets/slug URI, or a list of refs to fine-tune on sequentially."
    • addedInput schema / properties / epochs / description
      Added value: +"Maximum full passes over the training set."
    • addedInput schema / properties / imgsz / description
      Added value: +"Target input size: square batches normally, or the long-side size with rect=true."
    • addedInput schema / properties / name / description
      Added value: +"Run name for callbacks."
    • addedInput schema / properties / train_args / description
      Added value: +"Additional YOLO training arguments passed through to the platform. epochs, imgsz, batch, and name here are silently overridden by the matching top-level input when both are set; data and model are rejected outright if present here."
  2. Changed5 schema fields changedv0.1.12
    • changedInput schema / properties / confirm_cost / description
      Previous value: -"Must be true to allow a credit-costing training run."New value: +"Must be true to allow a credit-costing training run. Starting is billable immediately; the platform has no cost preview before that, so the estimated cost and remaining balance are only reported after the job starts."
    • addedInput schema / properties / confirm_history_loss
      Added value: +{
      +  "description": "Must be true to restart training on an existing model that already has a recorded run. Doing so replaces that model's status, epoch count, and per-epoch metric history irrecoverably; the previously uploaded weights survive. Not required for an untrained model or for checkpoint mode, which creates a new model instead. Separate from confirm_cost.",
      +  "type": "boolean"
      +}
    • addedInput schema / properties / dataset / anyOf
      Added value: +[
      +  {
      +    "type": "string"
      +  },
      +  {
      +    "items": {
      +      "type": "string"
      +    },
      +    "type": "array"
      +  }
      +]
    • changedInput schema / properties / dataset / description
      Previous value: -"Dataset ref used as training data for the job."New value: +"Dataset ref by slug, owner/slug, or ul:// URI, or a list of refs to fine-tune on sequentially."
    • removedInput schema / properties / dataset / type
      Removed value: -"string"
  3. Changed6 schema fields changedv0.1.6
    • addedInput schema / properties / confirm_cost / description
      Added value: +"Must be true to allow a credit-costing training run."
    • addedInput schema / properties / dataset / description
      Added value: +"Dataset ref used as training data for the job."
    • addedInput schema / properties / gpu_type / description
      Added value: +"Cloud GPU type to allocate for training."
    • addedInput schema / properties / model / description
      Added value: +"Existing model ref, or official YOLO base checkpoint such as yolo11n.pt or yolo11n-seg.pt. Checkpoint mode auto-creates a project model."
    • addedInput schema / properties / project / description
      Added value: +"Project ref that owns the training job and resolved model."
    • addedInput schema / properties / train_args
      Added value: +{
      +  "additionalProperties": {},
      +  "propertyNames": {
      +    "type": "string"
      +  },
      +  "type": "object"
      +}
  4. First observedv0.1.0

TDQS

A4.8/5.0
Behavior5/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations indicate destructiveHint=true and readOnlyHint=false, but the description goes far beyond that: it details exactly what is destroyed (status, epoch count, per-epoch metric history) and that it is irrecoverable; it discloses billing starts immediately with no cost preview; it reveals a side effect in checkpoint mode (leftover model) and names the delete tool. This is far richer than the annotations alone.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness4/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The description is long, but every sentence adds unique and essential information given the tool's complexity (multiple modes, destructive effects, cost). It is front-loaded with the purpose and then systematically explains edge cases. While slightly verbose, it avoids redundancy and is appropriately sized for the risk profile.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given the high complexity (11 parameters, nested objects, multiple modes), the description covers all calling requirements: prerequisites, side effects, confirmation flags, and failure modes. It does not describe the return value or output, which is a minor gap since no output schema is provided, but the agent has sufficient information to invoke and interpret the operation. Sibling tools like training_monitor partially fill the output gap.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters5/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema coverage is 100%, but the description adds critical parameter interplay: it clarifies that confirm_cost must be true, confirm_history_loss is required only for existing trained models, explains how train_args top-level conflicts are handled (silently overridden or rejected), and defines the semantics of dataset lists vs single dataset. This is substantial value beyond the schema.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

The description states a specific verb (start) and resource (cloud training job) and clearly distinguishes from siblings: it explicitly references training_cancel for stopping and mentions model_download, training_monitor, etc. indirectly. It differentiates from other training-related tools by describing its unique role as starting a job.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines5/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description explicitly states when to use the tool (to start training) and provides explicit conditions: it explains when confirm_history_loss is needed vs not, how checkpoint mode differs, and when cost confirmation is mandatory. It even names training_cancel as the alternative for stopping a running job, giving clear routing.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.