train
ManagementcatalogTrain Models
Not in the default tool list
This tool lives in the tool catalog. Discover it with scenario_tools_search (or enumerate everything with scenario_tools_list) and run it via scenario_tool_execute_write. To list every tool directly, connect with ?toolsets=full.
Manage custom model training. Dataset: upload_images, update_image, delete_image, set_pairs (replaces every pair: send all in one call). Runs: configure prices a config, start launches training, stop cancels it. config fields: epochs, nb_repeats, batch_size, learning_rate, rank, sample_prompts, sample_source_images, sample_width, sample_height, guidance_loss_scale, max_timestep, attention_only. Confirm the exact target before replacing, deleting or stopping.
action="start" is the ONLY action that launches training. It is paid and irreversible: CU are charged when the job is created. action="stop" refunds nothing, even at 0%. Get the user's explicit go-ahead on the quoted CU first.
action="configure" sends ?dryRun=true: no job, no charge. It prices the dataset as is, unvalidated: price after the last dataset change. To launch the quote, repeat as action="start" with the same config.
MiniMax H3 video LoRAs, by model_create data.type: minimax-h3-lora trains on images, clips or both; minimax-h3-int8-lora on images only, for less; minimax-h3-ref-lora on clips paired with their references. Flow: upload_asset, upload_images, set_pairs (Reference), captions, configure, start. Video entries bill far above images. Tell the user: Subject to the MiniMax H3 license (https://huggingface.co/MiniMaxAI/MiniMax-H3/blob/main/LICENSE). Its outputs may not be used to train other AI models.
Captions are asset descriptions; start refuses an uncaptioned entry. Clips get one 40 s to 2 min after upload_images, or none: check with asset_get, set it with asset_update metadata.description (catalog-only — run via scenario_tool_execute_write).
Returns: dataset changes, cost estimates or training state. Estimates carry creativeUnitsCost, training_started=false and no job. A real start returns a job; track it with jobs_wait or job_get.
Examples:
- "Start with 12 epochs" (quote approved) -> action="start", model_id="model_xxx", config={"epochs": 12, "sample_prompts": ["a stone bridge at night"]}
Don't use for inference — use model_run.
destructiveopen-world
Parameters
| Name | Type | Required | Description |
|---|---|---|---|
| action | enum(upload_images | update_image | delete_image | set_pairs | configure | start | stop) | ✓ | Only 'start' launches paid training. |
| model_id | string | ✓ | Model ID. |
| training_image_id | string | — | update_image/delete_image target (asset id). update_image swaps in a different asset: the training image takes its id (returned as trainingImage.id) and caption; the old caption does not carry over. |
| images | array | — | upload_images: upload_asset ids (not URLs), any count. Images; minimax-h3-lora/-ref-lora also mp4/webm/mov clips (5.25-30 s, ≤400 MB, only the first ~5.3 s trains); -ref-lora audio too. |
| image_data | record | — | update_image: assetId (existing asset) or data (data URL) plus name. Captions cannot be set here: see the tool description. set_pairs: { body: [{ sourceId, targetId, instruction }] }; H3 Reference, 5+ pairs: { body: [{ targetId, instruction, referenceIds }] }: a unique clip, its caption (else its description), 1-12 dataset assets in order (≤9 images, ≤3 videos of 2-15 s, ≤3 audio; ≥1 image or video). |
| config | unknown | — | Required by configure; pass the same to start. |
| dry_run | boolean | — | start only: true only estimates. configure always does; dry_run=false is rejected there. |
| task_id | string | — | Task of the team's task tracker (Jira, Wrike…) this call's Creative Units are attributed to. Required when the project enforces task ids: the call is then refused without one. Only an id the user chose from the project's task list or typed: never invent one, the API checks only its format. Pass the same task_id on later billed calls for the same work in the same project; ask again when the user moves to other work or switches project or team. A dry_run estimate needs none. project_tasks_list (catalog-only — run via scenario_tool_execute_read) lists the project's tasks. |
| team_id | string | — | Team ID. Required with project_id for OAuth callers; an API key is scoped by the key. |
| project_id | string | — | Project ID. Required with team_id for OAuth callers; an API key is scoped by the key. |
| response_format | enum(json | markdown) | json | Output format: 'json' for structured data, 'markdown' for human-readable text. |
Example Request
JSON
{
"action": "upload_images",
"model_id": "model_custom_abc",
"images": [
"asset_clip001",
"asset_clip002",
"asset_still003"
],
"team_id": "team_abc123",
"project_id": "proj_xyz789"
}Example Response
JSON
{
"trainingImages": [
{
"id": "asset_clip001",
"name": "run-cycle.mp4"
},
{
"id": "asset_clip002",
"name": "jump.mp4"
},
{
"id": "asset_still003",
"name": "pose.png"
}
]
}Common Use Cases
- Upload reference images to a model before starting fine-tuning
- Build a MiniMax H3 video LoRA dataset from images and 5.25-30 s video clips
- Price a configuration (epochs, learning rate, rank) before launching
- Start or stop a training job programmatically
- Pair before/after training images with an edit instruction for image-editing training