---

name: language-seq2seq description: "Plan and troubleshoot UniLM language sequence-to-sequence fine-tuning, decoding, evaluation, and related language adaptation workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

language-seq2seq

Use this sub-skill when a task asks for UniLM, UniLMv1, s2s-ft, MiniLM NLG, AdaLM, DeltaLM, InfoXLM, EdgeLM, LayoutReader, XTune, or legacy language sequence-to-sequence command construction and diagnosis.

What this covers

• Build safe command plans for summarization, question generation, JSONL seq2seq training, decoding, and ROUGE evaluation.
• Choose between legacy UniLMv1 scripts, newer s2s-ft scripts, MiniLM NLG paths, multilingual/domain-adaptation language components, and fairseq-style generation stacks.
• Validate src/tgt JSONL, tokenized list inputs, parallel train.src/train.tgt files, vocabulary/config/checkpoint expectations, and prediction/evaluation files.
• Diagnose old Torch/Apex/ROUGE/tokenizer/checkpoint failures without downloading model weights or starting heavy jobs.

First routing choice

• Use s2s-ft command patterns for modern UniLM/MiniLM/Roberta/XLM-R/Electra seq2seq JSONL workflows.
• Use UniLMv1 command patterns only when the request explicitly targets biunilm/run_seq2seq.py, biunilm/decode_seq2seq.py, legacy BERT vocab/config/checkpoints, or old train.src/train.tgt data.
• Use AdaLM notes for biomedical/computer-science domain-adapted BERT-style fine-tuning, not generic seq2seq generation.
• Use DeltaLM or EdgeLM notes for fairseq encoder-decoder translation/generation workflows.
• Route dense retrieval, OCR/form/image, multimodal generation, and architecture/distributed-research questions to sibling skills instead of this one.

Bundled references and scripts

• Read references/workflows.md to choose the correct UniLM-family language workflow and safety posture.
• Read references/api-and-cli.md for supported scripts, flags, model choices, and evaluation commands.
• Read references/data-formats.md before preparing JSONL, tokenized inputs, legacy parallel files, predictions, or ROUGE files.
• Read references/troubleshooting.md when dependencies, checkpoints, tokenizers, CUDA/Apex, distributed launch, or ROUGE evaluation fail.
• Run scripts/build_seq2seq_command.py to validate required planning arguments and print safe s2s-ft train/decode/eval command templates without executing them.
• Run scripts/inspect_legacy_tokenizers.py to inspect a local BERT/UniLM vocabulary and special-token compatibility without downloading models.

Safety rules

• Do not run training, decoding, distributed launch, downloads, ROUGE Perl, or GPU-only examples unless the user explicitly asks and the runtime is prepared.
• Treat UniLMv1 full scripts as reference-only unless an old PyTorch 1.1-style environment, matching Apex build, checkpoint, config, and vocab are confirmed.
• Prefer command planning and file validation first; then ask before executing any heavy or credentialed workflow.