---

name: pipelines-cli description: "Use Fengshen public pipeline APIs and the fengshen-pipeline console command for safe prediction/training orchestration, fixtures, and troubleshooting." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Fengshen Pipelines and CLI

Use this sub-skill when the task is to run, adapt, inspect, or troubleshoot Fengshen public pipeline surfaces or the fengshen-pipeline console command.

Start here

1. Read references/cli-reference.md before using fengshen-pipeline; the console command is generic, but not every pipeline class is console-compatible.
2. For classification prediction/training data, read references/text-classification.md and generate a tiny local fixture with scripts/make_classification_fixture.py.
3. For NER/sequence tagging, read references/sequence-tagging.md and generate the expected labels.txt plus *.all.bmes layout with scripts/make_sequence_tagging_fixture.py.
4. For UniMC, UniEX, and Ubert programmatic routes, read references/unimc-uniex-ubert.md.
5. For TCBert prompt classification, read references/tcbert.md.
6. When behavior is surprising, use references/troubleshooting.md before assuming the installed package is broken.

Safe checks first

These checks do not download models, start training, or mutate checkpoints:

python scripts/inspect_pipeline_cli.py --pipeline text_classification --pipeline sequence_tagging
python scripts/inspect_pipeline_cli.py --pipeline unknown_name
python scripts/make_classification_fixture.py --out-dir ./fengshen-classification-fixture
python scripts/make_sequence_tagging_fixture.py --out-dir ./fengshen-sequence-tagging-fixture

Use fengshen-pipeline text_classification predict --help only as a help/import check. Real predict and train calls may download model weights or datasets unless all paths are local and cached.

Route boundaries

• Model class selection, custom configs, tokenizers, and fengshen_model_type mapping belong to model-zoo.
• UniversalDataModule, checkpoints, optimizers, PyTorch Lightning Trainer flags, Deepspeed, CUDA, and low-level dataloaders belong to data-training.
• Large example command families, CLUE recipes, Stable Diffusion, Ziya, and checkpoint conversion planning belong to examples-conversion.
• This sub-skill owns pipeline entry points, parser shapes, task data schemas, safe fixtures, and CLI misuse diagnosis.

Do not do these from this sub-skill

• Do not run original repository tests or example shell scripts as a smoke check; several have hard-coded local paths or GPU assumptions.
• Do not treat --help success as proof that model downloads, CUDA execution, Deepspeed training, or checkpoint loading are verified.
• Do not use unsupported pipeline names or programmatic-only classes through fengshen-pipeline without first checking their route in references/cli-reference.md.