finetuning-data-prep
Prepare and validate MOSS-TTS fine-tuning JSONL data and launch supervised fine-tuning runs.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- finetuning-data-prep
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Prepare and validate MOSS-TTS fine-tuning JSONL data and launch supervised fine-tuning runs.
- disable-model-invocation
- true
02 资源文件 5
- 参考文档references/data-formats.md9.7 KB
- 参考文档references/preprocessing.md8.6 KB
- 参考文档references/training-launches.md7.9 KB
- 参考文档references/troubleshooting.md9.2 KB
- 脚本scripts/validate_training_jsonl.py24.2 KB
03 SKILL.md 正文
finetuning-data-prep
Use this sub-skill when the task is about preparing training JSONL, codec preprocessing, supervised fine-tuning launch planning, or first checkpoint smoke checks for MOSS-TTS Delay, Local, Local Transformer v1.5, Realtime, TTSD, SoundEffect v1, or VoiceGenerator workflows.
Use this for
- Choosing a raw JSONL schema for TTS, voice cloning, multi-speaker TTSD, SoundEffect v1, VoiceGenerator, or Realtime conversation fine-tuning.
- Validating raw or prepared JSONL before expensive codec preprocessing or
accelerate launchtraining. - Running or modifying the repository fine-tuning helpers:
prepare_data.py,sft.py,run_train.sh, Accelerate DDP/FSDP/DeepSpeed configs, and sharded JSONL outputs. - Planning batch size, gradient accumulation, mixed precision, loss weighting,
n_vq, and single-node or multi-node launch shapes. - Diagnosing data-preparation and training-launch failures before handing generation/inference work to an inference sub-skill.
Route elsewhere
- SoundEffect v2 DiT fine-tuning, diffusion-specific manifests, or DiT checkpoint issues:
../soundeffect-v2/SKILL.md. - llama.cpp conversion, quantization, backend evaluation, or GGUF runtime checks:
../llama-cpp-backend/SKILL.md. - Full inference after training belongs to the owning inference sub-skill for the selected model family. This sub-skill only covers a quick post-finetune smoke check that verifies a checkpoint can generate a short WAV.
- Generic Hugging Face model-family selection and non-training prompts:
../hf-family-workflows/SKILL.md.
Operating sequence
- Pick the task id and model family from
references/data-formats.md. - Validate the raw or prepared JSONL with
scripts/validate_training_jsonl.pybefore launching codec or model work. - Preprocess raw audio into
audio_codesor conversation-turn codes usingreferences/preprocessing.md. - Launch training with
references/training-launches.md; use the bundled troubleshooting reference if ranks, shards,n_vq, memory, or reference-audio fields fail. - After a checkpoint is saved, perform only a short smoke inference with the matching inference workflow, then route any deeper generation/evaluation request to the owning inference sub-skill.
Bundled references and tools
- JSONL schemas and task-specific required fields:
references/data-formats.md. - Codec preprocessing, reference-code handling, sharding, and validator usage:
references/preprocessing.md. accelerate, FSDP, DeepSpeed,run_train.sh, and hyperparameter launch patterns:references/training-launches.md.- Known failure modes and recovery actions:
references/troubleshooting.md. - No-model JSONL checker:
scripts/validate_training_jsonl.py.
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).