data-model-customization
Customize ms-swift datasets, registries, plugins, models, and templates safely.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- data-model-customization
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Customize ms-swift datasets, registries, plugins, models, and templates safely.
- disable-model-invocation
- true
02 资源文件 7
- 参考文档references/data-formats.md8.0 KB
- 参考文档references/registry-and-plugins.md8.8 KB
- 参考文档references/template-reference.md7.5 KB
- 参考文档references/troubleshooting.md7.8 KB
- 脚本scripts/inspect_registries.py11.7 KB
- 脚本scripts/inspect_template_encoding.py8.2 KB
- 脚本scripts/validate_dataset_rows.py24.0 KB
03 SKILL.md 正文
Data and Model Customization
Use this sub-skill when the task is to prepare or validate custom data, map columns, inspect ms-swift registries, register datasets/models/templates through an external plugin, or reason about model/template selection before training or inference.
Read First
- Data formats: standard
messagesrows,--columns, SFT/RLHF/multimodal/agent fields, and preprocessor selection. - Registry and plugins:
--custom_dataset_info,--external_plugins,register_dataset,register_model,register_template, and safe plugin structure. - Template reference:
get_processor,get_template, template/model selection, agent templates, loss scaling, and offline skeleton checks. - Troubleshooting: row drops, malformed messages, media problems, template mismatch, plugin side effects, and local/offline loading.
Bundled Checks
scripts/validate_dataset_rows.py: validate JSONL/JSON/CSV rows for common ms-swift keys before passing them toswift sft,swift pt, orswift rlhf.scripts/inspect_registries.py: list installedDATASET_MAPPING,MODEL_MAPPING, andTEMPLATE_MAPPINGentries without model downloads.scripts/inspect_template_encoding.py: print a safe template-encoding plan by default; optionally attempt an installed ms-swift processor/template encode without downloading weights.
Routing Boundaries
- Stay here for
--dataset,--columns,--custom_dataset_info,--external_plugins, preprocessors, registry inspection, and model/template registration plans. - Route full
swift sft,swift pt, and dataset packing/training execution to the training sub-skill. - Route inference engine behavior, deployment servers, and client requests to inference/deployment coverage.
- Route RLHF/GRPO algorithm execution, reward models, rollout workers, Ray, and Megatron execution to advanced distributed/RLHF coverage.
- Route export and evaluation execution to export/evaluation coverage.
Quick Patterns
- Prefer direct data files first:
swift sft --dataset train.jsonl --columns prompt=query answer=response .... - Use
--custom_dataset_info dataset_info.jsonwhen stable aliases, subsets, splits, or reusable column mappings are needed. - Use
--external_plugins plugin.pyonly when code must callregister_dataset,register_model, orregister_template. - Inspect support with
python scripts/inspect_registries.py --models --templates --contains qwen2_5before overriding--model_typeor--template. - Validate rows with
python scripts/validate_dataset_rows.py data.jsonl --columns raw_prompt=query raw_answer=response --check-media existsbefore long runs.
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).