Lora Training

Train and retrieve a LoRA with RunComfy

Use the connected RunComfy MCP tools. If disconnected, use the client's authentication flow; do not request tokens in chat. Explaining LoRA settings alone does not require creating a dataset or starting a job.

Prepare a reproducible configuration

1. Establish the user's training purpose, base model, dataset, desired output and spending limit. Verify supported base-model/trainer settings in current RunComfy documentation or the user's known working AI Toolkit config. An inference model's availability does not imply that it supports training.
2. Use list_datasets to find the intended dataset. For a new dataset, create and upload files only within the user's authorization. Preserve exact filenames and matching caption names.
3. upload_dataset_file_from_url imports media from an accessible URL; upload_dataset_text_file writes caption text. For local files, get_dataset_upload_urls provides signed upload destinations. Upload through a supported client tool without exposing signed URLs or credentials in prose. If no upload tool is available, explain that limitation.
4. Verify the resulting file inventory and get_dataset_status readiness before submission. Do not treat upload acceptance as a READY dataset.
5. Prepare the complete AI Toolkit YAML from verified configuration, including model/adapter compatibility, steps, learning rate, rank, batch size, resolution, checkpoint interval and sample settings. Use a unique job name consistently across the YAML. Do not silently rename an existing job or resume it.
6. Respect the platform path contract: training_folder is /app/ai-toolkit/output; dataset folder_path is /app/ai-toolkit/datasets/{dataset_name} using its name, not ID. These are remote container paths, not files to create on the user's computer.
7. Check current GPU options/rates and get_balance. Show the exact YAML, GPU type/count, dataset, estimated cost, spending limit and monitoring/stop conditions. Proceed only with the user's approval of that concrete run; previously approved unchanged settings do not need repeated confirmation.

Submit, monitor and report

• Call submit_training_job once with the unchanged YAML and approved GPU settings. Record its job ID immediately.
• Monitor get_training_job_status with reasonable spacing. A timeout or client interruption does not justify another submission. Reconcile the existing job first.
• Use cancel_training_job only for user cancellation or an agreed stopping condition. resume_training_job, edit_training_job, dataset deletion and additional jobs require authorization appropriate to their effects.
• Retrieve get_training_job_result and report actual checkpoint/config/sample artifacts. A terminal STOPPED state alone does not prove successful training: inspect completed steps, errors and available artifacts.
• When the task includes stopping GPU charges, verify GPU release from available status or the RunComfy job page; distinguish that from process completion. Report an unavailable release signal honestly.
• Inspect samples when supported. A downloaded checkpoint or completed training job is not a separate inference/reload test or proof of output quality.
• If the user wants to run the LoRA, verify a compatible base model and its current get_model input schema, then obtain approval for that additional inference cost. Never assume every model accepts the same LoRA field.

Useful requests

• “Inspect my dataset and review a short LoRA training configuration and GPU cost. Do not start yet.”
• “Check this training job's progress and return its checkpoints and final sample without resuming it.”

AI Toolkit LoRA training · Trainer API documentation