data-preparation
Prepare and validate WeNet data manifests, dictionaries, tokenizers, CMVN inputs, and raw-versus-shard dataset layouts before training or decoding.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-preparation
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Prepare and validate WeNet data manifests, dictionaries, tokenizers, CMVN inputs, and raw-versus-shard dataset layouts before training or decoding.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/data-formats.md3.8 KB
- 参考文档references/troubleshooting.md3.1 KB
- 参考文档references/workflows.md3.2 KB
- 脚本scripts/make_wenet_dict.py2.7 KB
- 脚本scripts/make_wenet_raw_manifest.py3.3 KB
03 SKILL.md 正文
WeNet Data Preparation
Use this sub-skill when the task is to convert speech data into WeNet training
or evaluation inputs, validate wav.scp/text, create data.list, build a
character dictionary, choose raw versus shard data, or debug tokenizer/CMVN
setup.
Start here
-
Read references/data-formats.md to confirm the expected
wav.scp,text,data.list, dictionary, and tokenizer shapes. -
Use the bundled raw-manifest helper for small or medium custom datasets:
python sub-skills/data-preparation/scripts/make_wenet_raw_manifest.py \ --wav-scp wav.scp --text text --output data.list -
Use the bundled dictionary helper for character-token recipes:
python sub-skills/data-preparation/scripts/make_wenet_dict.py \ --text text --output units.txt -
Read references/workflows.md before planning CMVN, shard data, tokenizer selection, or recipe-stage conversion.
-
Read references/troubleshooting.md when dataset loading filters everything, keys mismatch, audio cannot be decoded, dictionaries have wrong reserved IDs, or tokenizer files cannot be found.
Route by task
- Single-audio package transcription belongs in ../package-transcription/SKILL.md.
- Training, checkpoint averaging, recognition, and WER/CER scoring belong in ../training-and-decoding/SKILL.md.
- Exporting a trained checkpoint belongs in ../model-export/SKILL.md.
Key decisions
- Choose
rawdata for ordinary small/medium datasets and custom debugging. Chooseshardonly when the dataset is large enough that tar shards improve throughput and the user can tolerate extra packaging steps. - Use absolute or stable audio paths in
wav.scpanddata.list. Relative paths often fail when training is launched from a different directory. - For Mandarin character recipes, dictionary files reserve
<blank> 0,<unk> 1, and<sos/eos>as the final or recipe-defined special token. Keep the active training config and tokenizer initialization consistent. - Treat CMVN computation and feature extraction as data-prep stages, but do not run large audio scans unless the user has approved the data path, runtime, and storage cost.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).