AGENT SKILL

synthetic-data-generator

Use SDGX/Synthetic Data Generator for tabular synthetic data, metadata inspection, relationship metadata, CLI fit/sample, LLM synthesis, evaluation metrics, and plugin-oriented extension workflows.

License 许可
Apache-2.0
最近更新
2026-09-03
Stars
★ 351
Forks
27

01 前言 Front Matter

name
synthetic-data-generator
license
Apache 2.0
metadata
{ "disco-role": "operating" }
description
Use SDGX/Synthetic Data Generator for tabular synthetic data, metadata inspection, relationship metadata, CLI fit/sample, LLM synthesis, evaluation metrics, and plugin-oriented extension workflows.
disable-model-invocation
true

02 资源文件 6

03 SKILL.md 正文

Synthetic Data Generator (SDGX) Repo Skill

Use this skill when a task names sdgx, Synthetic Data Generator, or asks for structured/tabular synthetic data generation with SDGX APIs or the sdgx CLI. The package centers on pandas tables, metadata inspection, default data processors, CTGAN/GaussianCopula synthesis, OpenAI-backed tabular generation, and tabular quality metrics.

Do not use this skill for generic GAN theory, unrelated synthetic image/text generation, or privacy-policy decisions that are not about operating SDGX APIs.

Fast orientation

  • Public package/distribution: sdgx.
  • Main import check: import sdgx; print(sdgx.__version__).
  • Main CLI: sdgx with fit, sample, and list-* subcommands.
  • Primary high-level API: sdgx.synthesizer.Synthesizer.
  • Canonical single-table ML model: sdgx.models.ml.single_table.ctgan.CTGANSynthesizerModel.
  • Statistic model available by direct import: sdgx.models.statistics.single_table.copula.GaussianCopulaSynthesizerModel.
  • LLM model: sdgx.models.LLM.single_table.gpt.SingleTableGPTModel.

Read references/install-and-import.md before installing, verifying imports, or choosing CPU/CUDA/OpenAI settings. Run scripts/check_sdgx_environment.py when you need a quick registry/import/backend check in the active environment.

Quick install

pip install sdgx
python - <<'PY'
import sdgx
print(sdgx.__version__)
PY
sdgx --help

For local checkout work, pip install . or pip install '.[test]' from the repository root are the documented alternatives.

Route by task

  • Data connectors, loaders, metadata, processors, cachers, inspectors, relationships, and plugins: read sub-skills/data-preparation/SKILL.md. Start here when the task is about CSV/DataFrame/generator inputs, metadata JSON, datetimes, fixed/specific column combinations, PII columns, caching, or custom extension registration.
  • Fit/sample tabular synthesizers or use the sdgx CLI: read sub-skills/single-table-synthesis/SKILL.md. Start here for Synthesizer, CTGAN, GaussianCopula, model save/load, sdgx fit, sdgx sample, --json_output, and --torchrun.
  • Generate tabular data with OpenAI-compatible LLMs or infer off-table features: read sub-skills/llm-synthesis/SKILL.md. Start here for SingleTableGPTModel, OPENAI_KEY, OPENAI_URL, metadata-only generation, raw-data prompting, and response parsing.
  • Evaluate real vs synthetic data or adapt benchmark-style checks: read sub-skills/evaluation/SKILL.md. Start here for Jensen-Shannon divergence (JSD), mutual-information similarity (MISim), and safe metric scripts.

Common operating pattern

  1. Choose a data connector or create a pandas.DataFrame.
  2. Build or inspect Metadata and optionally customize column types, primary keys, specific combinations, categorical encoders, and datetime formats.
  3. Choose a model and entrypoint:
    • Library: Synthesizer(model=..., data_connector=..., metadata=...) then fit() and sample(n).
    • CLI: sdgx fit ... then sdgx sample ....
    • LLM: SingleTableGPTModel with raw data or metadata, after confirming key/base URL policy.
  4. Validate outputs: expected columns, row count, null/datetime/PII behavior, and metric ranges.
  5. Use troubleshooting references when imports, metadata checks, cache writes, optional dependencies, CUDA, or API credentials fail.

Repo-level references

Guardrails

  • Never assume the package has an installed inspection environment just because the checkout imports from the current directory; verify with the target Python.
  • Do not run notebook or benchmark-scale workflows as quick checks. Use tiny fixtures and short smoke scripts unless the user explicitly requests benchmark execution.
  • Treat LLM generation as a network/credential workflow. Do not send sensitive raw rows to an external API unless the user explicitly authorizes that data flow.
  • When debugging an SDGX error, preserve the exact exception class from sdgx.exceptions; CLI --json_output true can surface structured exit messages.

04 同仓库更多 Skills

VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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.

★ 351
VectorSpaceLab
VectorSpaceLab/AREX-Skill

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).

★ 351