training

Plugin training from marketplace aiwg in jmagly/aiwg. Commands read from the repo's manifest — today.

Skills (15)

• acquire-training-source

  Acquire a training data source with license validation and delegate ingest to the semantic memory kernel

• dataset-docs

  Generate Datasheet, Model Card, and Data Statement from a dataset manifest

• dataset-reproduce

  Deterministically rebuild a dataset from its manifest and verify fixity equivalence

• dataset-version

  Create a versioned training dataset with manifest, fixity, provenance, and archive snapshot

• decontamination-check

  Detect training-eval overlap against benchmark sets before dataset publication

• example-quality-assess

  GRADE quality assessment adapted for individual training examples

• example-synthesizer

  Generate SFT training examples from raw sources using Self-Instruct / Evol-Instruct / SQuAD / STaR patterns

• flow-dataset-build

  End-to-end training dataset pipeline — acquire sources through publication

• format-adapter-alpaca

  Convert canonical training examples to Alpaca format for training frameworks

• format-adapter-chatml

  Convert canonical training examples to ChatML format for training frameworks

• format-adapter-jsonl

  Convert canonical training examples to JSONL format for training frameworks

• format-adapter-parquet

  Convert canonical training examples to Parquet format for training frameworks

• format-adapter-sharegpt

  Convert canonical training examples to ShareGPT format for training frameworks

• preference-generator

  Generate preference pairs (chosen/rejected) for DPO/KTO/ORPO/SimPO training

• synthetic-data-generator

  Large-scale synthetic training data generation with Model Collapse recursion guard

Subagents (7)

• dataset-evaluator-agent

  Computes dataset-level metrics (diversity, difficulty, domain balance, quality grade distribution) and prepares the matric-eval handoff package for model evaluation.

• dataset-publication-agent

  Coordinates dataset versioning, datasheet/model card generation, integrity manifests, and the publication gate including override escalation paths.

• decontamination-agent

  Runs exact, fuzzy, and semantic contamination checks against eval-set targets and feeds the publication gate.

• example-synthesizer-agent

  Generates SFT training examples from admitted sources using self-instruct, evol-instruct, squad, and STaR patterns with per-example provenance.

• format-converter-agent

  Runs mechanical format adapters (alpaca, sharegpt, chatml, jsonl, parquet) with round-trip validation and sidecar metadata.

• preference-generator-agent

  Generates DPO/KTO preference pairs from candidate examples using llm-judge, rule-based, or human modes, with calibrated confidence and edge-of-graph writes.

• source-curator-agent

  Decides which sources are admitted to the training corpus and drives license, format, and quality gating at ingest.

Similar plugins