---

name: ofa description: "Routes OFA workflows for multimodal pretraining, captioning, VQA, RefCOCO, OCR, ImageNet, image generation, Gigaword, GLUE, and MMSpeech, with safe setup and validation helpers." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

OFA

OFA is a unified sequence-to-sequence repository for vision-language, text, image generation, language understanding, and speech workflows.

Use this skill when a user wants to:

• run or adapt OFA training or evaluation commands,
• validate OFA TSV / manifest layouts before a long GPU job,
• reproduce caption, VQA, RefCOCO, OCR, ImageNet, Gigaword, GLUE, or MMSpeech workflows,
• inspect OFA model, task, or criterion registration,
• understand prompt tuning, adapters, bitfit, or encouraging loss,
• or reason about the repo's bundled Fairseq fork and command-line entry points.

Start here

1. Read references/repo-provenance.md if you need to know whether this skill matches the current checkout.
2. Run scripts/check_ofa_environment.py with --check-clis to confirm the repo import path, CUDA availability, and CLI help surface.
3. Use scripts/render_ofa_command.py when you need a copyable launch command instead of hand-writing a long distributed invocation.

Shared helpers

• scripts/check_ofa_environment.py checks imports, package versions, CUDA readiness, and optional CLI help.
• scripts/render_ofa_command.py renders a safe OFA train/evaluate command without executing it.

Router map

setup-and-command-building

Use this for installation, PYTHONPATH, --user-dir, distributed launch shape, port selection, and command rendering. It is the first stop when a user says "how do I run OFA" or "why does train.py import fail?"

data-formats

Use this for TSV rows, selected-column layouts, base64 image payloads, image-code sequences, COCO-style caption rows, VQA answer fields, RefCOCO boxes, OCR data, ImageNet label rows, Gigaword/GLUE tables, and MMSpeech manifests.

pretraining

Use this for multimodal pretraining and continuous pretraining, including the vision_language_examples.tsv, text_examples.tsv, image_examples.tsv, detection_examples.tsv, and negative_sample/ layout.

vision-language-tasks

Use this for captioning, VQA, RefCOCO / RefCOCO+ / RefCOCOg, SNLI-VE, OCR, and ImageNet finetuning or evaluation.

image-generation

Use this for text-to-image generation, VQGAN code conversion, CLIP ranking, and image-generation validation.

language-tasks

Use this for Gigaword summarization and GLUE-style text understanding tasks such as CoLA, MNLI, RTE, QNLI, QQP, MRPC, and SST-2.

mmspeech

Use this for MMSpeech staged ASR pretraining/evaluation, speech manifests, fbank configuration, and WER troubleshooting.

model-internals-and-extension

Use this for task/model/criterion registration, architecture inspection, prompt tuning, adapters, bitfit, encouraging loss, and adding new OFA extensions.

Installation

From the repo root, install the public dependencies with:

pip install -r requirements.txt

Public prerequisites

• The repo's Python requirements.
• A CUDA-capable environment for the task workflows; CPU is only a partial substitute for setup checks.
• Java 1.8 if you want full COCO caption evaluation with SPICE.
• External datasets and checkpoints for the actual task runs.

Minimal verification

After install, run:

python scripts/check_ofa_environment.py --check-clis

A healthy result shows the repo imports, the bundled Fairseq fork is visible, CUDA is available when expected, and train.py / evaluate.py print help successfully.

Read more

• references/troubleshooting.md for cross-cutting import, backend, Java, and dependency issues.
• references/repo-provenance.md for the source commit, branch, and staleness baseline.
• references/repo-routing-metadata.json for router placement and scenario metadata.