Skills

Skill 26 of 34

Lora Training

Prepare datasets and manage AI Toolkit LoRA training on RunComfy GPUs.

2 minutes · 492 words · 4 sections

Install

$ npx skills add runcomfy-com/skills --skill lora-training

$ npx skills add runcomfy-com/skills

» /plugin marketplace add runcomfy-com/skills

The first command installs just this skill, by the name in its SKILL.md; the second installs the whole repository.

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

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.
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.
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.
Verify the resulting file inventory and get_dataset_status readiness before submission. Do not treat upload acceptance as a READY dataset.
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.
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.
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 (opens in a new tab) · Trainer API documentation (opens in a new tab)

About this skill

Trigger

Prepare datasets and manage AI Toolkit LoRA training on RunComfy GPUs. Use for reviewing training YAML, captions, dataset readiness, starting an approved job, monitoring progress and retrieving checkpoints or samples. Verify base-model compatibility and costs before training.

The verbatim description from this skill’s front matter — the string an agent matches on to decide whether to load it.

Colophon for this issue

Colophon

How this issue was made

Source

github.com/runcomfy-com/skills, branch main, last pushed 15 September 2026.

Licence

MIT — the text of every skill is reproduced unmodified, frontmatter included, under the upstream licence.

Discovery

34 skills found by walking the repository tree for SKILL.md, not by matching a directory convention. 2 distinct layouts observed: */SKILL.md, plugins/runcomfy/skills/*/SKILL.md.

Issue colours

Resolved from a deterministic hash of the owner name. Two accent tones are generated per issue and each is proven against its own ground before it ships: a single accent that passes AA on both light and dark paper is arithmetically impossible.

Heading repairs

1 repair applied to this skill so the document has one h1 and no skipped levels:

Shifted “4 headings” from h1 to h2 so the skill title is the only h1.

Spec compliance

2 editorial notes across 2 of 34 skills. They are printed in the margin of each skill rather than as errors here.

AI Music — 1 note
Runcomfy CLI — 1 note

Images

Images inside a skill come from the upstream repository. Where the author gave no alternative text we mark the image decorative rather than inventing a description — a plausible caption we made up is worse than none for the reader who depends on it.

Marketplace

A plugin manifest is published at .claude-plugin/marketplace.json by RunComfy, declaring 1 plugin. It is read for editorial metadata only — never as the skill index, which is always the repository tree.

Signal

This repository is unranked on skills.sh, which says nothing about it: the leaderboard covers a fraction of published skills.

Agent surfaces

The whole issue is available as one markdown document at /runcomfy-com/skills.md, and each skill at its own .md URL.

Publication

Set by Skills Docs from the source repository. Body text is Literata at the reader’s chosen size and measure; code is Geist Mono. Nothing on this page was written by us except this paragraph.

Train and retrieve a LoRA with RunComfy

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.

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.

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.

Verify the resulting file inventory and get_dataset_status readiness before submission. Do not treat upload acceptance as a READY dataset.

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.

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.

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.

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.

“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.”