---

name: language-tasks description: "Guides OFA Gigaword summarization and GLUE-style language understanding workflows." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

language-tasks

Use this sub-skill when a user wants to run OFA's language-only or mostly language workflows: Gigaword summarization or GLUE-style classification framed as seq2seq generation.

Trigger phrases

• "Run Gigaword"
• "ROUGE evaluation for OFA"
• "Train CoLA / MNLI / QQP / RTE / SST-2 with OFA"
• "Why does datasets.load_metric fail?"
• "What selected columns do the GLUE tasks use?"

What this sub-skill owns

• Gigaword training and ROUGE evaluation,
• GLUE task families (CoLA, MNLI, MRPC, QNLI, QQP, RTE, SST-2),
• prompt-type guidance for classification-as-generation,
• ROUGE JSON helper usage,
• task-specific selected columns and label mappings.

What it excludes

• visual entailment -> vision-language-tasks,
• pretraining -> pretraining,
• image-generation workflows -> image-generation,
• speech workflows -> mmspeech.

Read these files

• references/workflows.md for the Gigaword and GLUE workflow shapes.
• references/troubleshooting.md for metric and data-layout failures.
• scripts/eval_rouge_json.py for offline ROUGE evaluation from prediction JSON.

Typical workflow

1. Identify the GLUE task or summarization task.
2. Confirm the selected columns and split names.
3. Decide whether the task is evaluation-only or includes finetuning.
4. Use the bundled ROUGE helper when you need a standalone metric check.

Notes

• Gigaword uses a plain source/target pair.
• GLUE tasks are represented as seq2seq classification with yes/no/maybe style outputs.
• The same base OFA architecture is reused across these tasks; the data layout is what changes most.