name	flow-dataset-build
description	End-to-end training dataset pipeline — acquire sources through publication
namespace	training-complete
category	flow
platforms	["claude","copilot","cursor","factory","windsurf","warp","codex","opencode","openclaw","hermes"]
commandHint	{"argumentHint":"<config-file> [--stages <stage1,stage2,...>] [--dry-run] [--version <v>] [--interactive]"}

Flow: Dataset Build

End-to-end orchestrator that runs the full corpus-to-dataset pipeline in a single invocation, chaining every downstream training-complete skill from acquisition through published dataset version.

When to Use

You have a pipeline config file and want a complete dataset with one command
You are re-running a known-good pipeline against a new source manifest or a new version
You need a reproducible, auditable build with per-stage logs and a pipeline-level report
You want human authorization gates before expensive/irreversible stages

Do NOT use this flow for ad hoc experiments; invoke individual stage skills directly when iterating on a single stage.

Parameters

<config-file> — path to pipeline config YAML specifying sources, synthesis patterns, preference mode, format targets, decontamination targets (see schema below)
--stages <list> — comma-separated subset of stages to run (default: all). Example: --stages acquire,quality-assess
--dry-run — simulate every stage; validate config, print intended actions, write no artifacts
--version <v> — target version string for published dataset (overrides config.version_pattern default)
--interactive — pause at every stage boundary for human approval (per @native-ux-tools rule)
--continue-on-warn — do not block on WARNING-level lint findings (default: strict; ERROR always blocks)
--acknowledge-license-risk — bypass license-check gate on ERROR (requires explicit acknowledgement in pipeline report)
--acknowledge-contamination — bypass decontamination gate on ERROR (same requirement)

Default Pipeline (10 Stages)

acquire — run acquire-training-source once per source declared in config.sources. Writes raw corpus to .aiwg/training/working/<run-id>/raw/.
quality-assess — run example-quality-assess against sources and raw examples. Emits GRADE-style quality scores per source and per example.
license-check — lint gate. Block on ERROR-level license findings unless --acknowledge-license-risk supplied. Incompatible licenses fail the pipeline here.
synthesize — run example-synthesizer if config.synthesis declares patterns. Optional; skipped silently if absent.
synthetic-bulk — run synthetic-data-generator if config.synthetic_generator_config present. Optional.
preference — run preference-generator if config.preference_generation declares a mode (DPO/RLHF/constitutional). Optional.
format — run format adapters per config.format_exports: any of alpaca, sharegpt, chatml, jsonl, parquet. Each adapter is a separate skill.
decontamination — run decontamination-check against config.decontamination_targets plus default targets (MMLU, HumanEval, GSM8K, HellaSwag, TruthfulQA, ARC, Winogrande).
decontamination-gate — lint gate. Block on ERROR-level contamination overlap unless --acknowledge-contamination supplied.
publish — run dataset-version. Creates manifest, SHA-256 fixity, W3C PROV provenance record, and archive snapshot at datasets/<version>.yaml.

Stage Selection

--stages acquire,quality-assess runs only those two stages.
Prerequisite stages skipped explicitly emit a WARNING (e.g., running format without quality-assess is permitted but flagged).
Config may declare skip_stages: [synthetic-bulk] to default-skip a stage every run.
Stage dependencies (informational, not enforced): acquire → quality-assess → license-check → {synthesize, synthetic-bulk, preference} → format → decontamination → decontamination-gate → publish.

Human Authorization Gates

Per @human-authorization rule, the pipeline pauses for explicit human approval at these points:

Between stages 3 and 4 — after license-check passes, before invoking synthesis. Synthesis is expensive and rework is painful; confirm license posture first.
Between stages 9 and 10 — after decontamination passes, before publishing. Publishing creates an immutable versioned artifact.
Every stage boundary — when --interactive is set.

Gates use the platform-native question tool when available (AskUserQuestion on Claude Code; fallback to formatted stdout elsewhere).

Config File Schema

# pipeline-config.yaml
version_pattern: "v{major}.{minor}.{patch}"   # overridden by --version
split_ratios:
  train: 0.9
  validation: 0.05
  test: 0.05

sources:
  - uri: "hf://datasets/example/source1"
    license: "apache-2.0"
  - uri: "https://example.com/corpus.jsonl"
    license: "cc-by-4.0"

license_policy:
  allowlist: ["apache-2.0", "mit", "cc-by-4.0", "cc0-1.0"]
  blocklist: ["cc-by-nc-4.0", "proprietary"]

synthesis:                        # optional — omit to skip stage 4
  patterns: ["qa-rewrite", "chain-of-thought"]
  max_examples: 5000

synthetic_generator_config:       # optional — omit to skip stage 5
  backend: "openai:gpt-4o"
  target_count: 10000

preference_generation:            # optional — omit to skip stage 6
  mode: "dpo"                     # dpo | rlhf | constitutional
  rater_model: "claude-opus-4"

format_exports: ["alpaca", "jsonl", "parquet"]

decontamination_targets:
  - "mmlu"
  - "humaneval"
  - "custom:./eval/internal-holdout.jsonl"

skip_stages: []                   # e.g., ["synthetic-bulk"] to default-skip

Error Handling

Any stage error aborts the pipeline immediately.
--continue-on-warn downgrades WARNING-level lint findings to informational; ERRORs still abort.
Partial outputs are preserved in .aiwg/training/working/<run-id>/ for post-mortem and resumption.
Resumption is manual: re-invoke with --stages <remaining-stages> pointing at the same run-id.

Output

Per-stage logs — every stage appends structured events via memory-log-append to the run-scoped log at .aiwg/training/working/<run-id>/events.jsonl.
Pipeline-level report — human-readable Markdown at .aiwg/training/reports/pipeline-<version>-<timestamp>.md summarizing each stage, gate decisions, authorization records, and final artifact pointers.
Final artifact — datasets/<version>.yaml manifest plus sibling outputs (provenance, fixity, archive snapshot) produced by dataset-version.
Activity log entry — per @activity-log rule, one line appended to .aiwg/activity.log.

Examples

# Full pipeline
aiwg flow-dataset-build ./configs/instruct-v3.yaml --version v3.1.0

# Subset: acquire and quality-assess only (dry run iteration)
aiwg flow-dataset-build ./configs/instruct-v3.yaml \
  --stages acquire,quality-assess

# Dry-run the whole pipeline to validate config before committing compute
aiwg flow-dataset-build ./configs/instruct-v3.yaml \
  --dry-run --version v3.1.0-rc.1

Delegation

Downstream skills invoked by this flow:

acquire-training-source — stage 1
example-quality-assess — stage 2
example-synthesizer — stage 4
synthetic-data-generator — stage 5
preference-generator — stage 6
format-adapter-alpaca, format-adapter-sharegpt, format-adapter-chatml, format-adapter-jsonl, format-adapter-parquet — stage 7
decontamination-check — stage 8
dataset-version — stage 10

References

ADR-022 Section D2 — training dataset pipeline architecture
@human-authorization rule — authorization gate requirements
@native-ux-tools rule — interactive prompt patterns
@activity-log rule — post-run logging requirement

flow-dataset-build

Plus depuis ce dépôt