ModelScope Local Vision Trainer

Use this skill when the user wants local vision training on a Linux GPU workstation or server and describes the job in natural language.

Request Style

• Accept requests such as:
  • Train an image classifier locally with ./data/train and ./data/val, epoch=5.
  • 我想在本地做目标检测，训练标注是 ./data/annotations/train.json，验证标注是 ./data/annotations/val.json。
• Infer missing fields where that is safe.
• Default task: classification
• Default model:
  • Classification: google/vit-base-patch16-224
  • Detection: facebook/detr-resnet-50
  • Segmentation: facebook/sam-vit-base

What You Must Create

Run scripts/prepare_local_vision_training_workspace.py with the user's exact request text. It must create:

1. configs/local.vision.training.json
2. configs/local.vision.required.env
3. configs/local.vision.train.env
4. scripts/run_local_vision_training.sh

Execution Rule

• This skill always pauses once before execution.
• First create configs/local.vision.required.env with non-empty default values.
• Tell the user to review or edit that file.
• Tell the user to reply with exactly continue.
• Only after the user replies continue, run bash scripts/run_local_vision_training.sh.
• On macOS or any unknown host, stop after writing the files and report that execution should happen on the intended Linux GPU host.

Primary Script

• scripts/prepare_local_vision_training_workspace.py

Workflow

1. Read the user request and infer the task, model, dataset paths, and training parameters.
2. Run scripts/prepare_local_vision_training_workspace.py --request "<user text>".
3. Inspect the generated plan, configs/local.vision.required.env, and launch script.
4. Ask the user to review or edit configs/local.vision.required.env.
5. Tell the user to reply with exactly continue.
6. Only after continue, run bash scripts/run_local_vision_training.sh.

Task Rules

• Classification keywords: classification, 分类
• Detection keywords: detection, 目标检测
• Segmentation keywords: segmentation, segment, 分割, sam
• When task is ambiguous, default to classification

Command Rules

• Keep the generated launch script readable and editable.
• configs/local.vision.required.env should start with sensible defaults instead of blank placeholders.
• The launch script must verify that the configured local training entrypoint exists before it tries to run.
• Prefer one process by default unless the user clearly asks for distributed local training.

References

• references/runtime_requirements.md

Guardrails

• Never skip the continue gate.
• Never auto-run on macOS or an unknown host.
• Never assume the local training entrypoint already exists without checking.
• Never claim built-in local execution for a task unless the configured entrypoint is present on the target machine.