---

name: training-and-decoding description: "Run and adapt WeNet training recipes, checkpoint averaging, offline decoding, recognition modes, scoring, LM or k2 decoding, and training failure recovery." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

WeNet Training and Decoding

Use this sub-skill when the user wants to train or fine-tune a WeNet model, resume an experiment, run offline recognition on a data.list, average checkpoints, compare decoding modes, compute WER/CER, or reason about recipe stages.

Start here

1. Ensure data manifests, dictionaries, and tokenizer resources are valid with ../data-preparation/SKILL.md.

2. Read references/workflows.md for the staged training/decoding recipe pattern and large-job boundaries.

3. Read references/cli-reference.md for the installed-module command shapes and important options.

4. Use scripts/score_text.py for small local WER/CER checks when the full recipe scorer is unavailable:

   python sub-skills/training-and-decoding/scripts/score_text.py \
     --reference ref.txt --hypothesis hyp.txt --unit word
   

5. Read references/troubleshooting.md when distributed training hangs, NCCL/HCCL/gloo fails, DeepSpeed configs disagree with WeNet configs, checkpoints are missing, or decoding output is empty.

Route by task

• Use this sub-skill for python -m wenet.bin.train, recognize, average_model, and alignment workflows.
• Use ../model-export/SKILL.md after a trained checkpoint and train.yaml exist.
• Use ../runtime-deployment/SKILL.md after export when the user needs C++/mobile/web/server deployment.
• Use ../package-transcription/SKILL.md for one-off package CLI transcription of a single audio file.

Verification boundary

Real training, large decoding, LM graph creation, and k2/LF-MMI workflows can require large corpora, GPUs, distributed launchers, external toolchains, and long runtimes. Treat them as documented operational workflows unless the user explicitly authorizes the data, hardware, and time budget.