llm_dataset

Null Safety Dart CI Codecov GitHub Tag New Commits Last Commits Pull Requests Code size License

A stream-oriented Dart package for building, storing, querying, and consuming LLM training datasets. Provider-independent infrastructure — not a training engine. You own the sources, generators, validators, and trainer; this package wires the pipeline and keeps every entry versioned, queryable, and reproducible.

source → generate → create variations → validate → store → query → stream into trainer

import 'package:llm_dataset/llm_dataset.dart';

Future<void> main() async {
  final store = MemoryDatasetStore();
  final config = GeneratorConfig(
    dataset: 'general',
    datasetVersion: 'v1',
    language: 'en',
    pipelineVersion: 'pipe-1',
    seed: 42,
  );

  final pipeline = DatasetPipeline(
    source: MemorySource([
      DatasetSourceDocument(
        id: 'france',
        content: 'Paris is the capital of France.',
        title: 'France',
        language: 'en',
      ),
    ]),
    generator: QuestionAnswerGenerator(config),
    variations: [
      RuleBasedVariationGenerator(
        strategies: [
          VariationStrategy.paraphrase,
          VariationStrategy.formal,
          VariationStrategy.casual,
        ],
        seed: 42,
      ),
    ],
    validators: [
      EmptyContentValidator(),
      LengthValidator(minInputLength: 3),
      DuplicateValidator(store: store),
    ],
    store: store,
    dataset: 'general',
    datasetVersion: 'v1',
  );

  final result = await pipeline.run();
  print('stored=${result.entriesStored} rejected=${result.entriesRejected}');

  final dataset = Dataset(
    store: store,
    variationSelection: VariationSelection.onePerGroup,
  );
  await for (final batch in dataset.batches(8, shuffle: true, seed: 7)) {
    // trainer.train(batch);
    print('batch ${batch.length}');
  }
}

DatasetEntry is the unit of training data. Variations of the same semantic example share a variationGroup; canonical rows use variationIndex: 0 with no provenance.parentEntryId. Tool-calling / chat turns live under metadata['messages'] with DatasetToolCall objects embedded in assistant turns.

API Documentation

Generate docs locally with dart doc, then open doc/api/index.html. The public surface is exported from lib/llm_dataset.dart.

Features

  • Pipeline-first. DatasetPipeline streams source documents through a generator, variation generators, validators, and into a store — rejecting bad rows without stopping the run.
  • Variation groups. Canonical entries plus indexed surface-form variations share variationGroup; provenance links variations back to their parent.
  • Built-in generators. Rule-based TextGenerator, QuestionAnswerGenerator, SummarizationGenerator, TranslationGenerator, ClassificationGenerator, ExtractionGenerator, and ToolCallGenerator — no LLM SDK required.
  • Translation variations. TranslationVariationGenerator with callback or rule-based backends; merge and uniquify multiple target languages in one pass.
  • Text exercises. EN/PT phrase and paragraph generators, JSON-backed phrase template stores, combinatorial expansion, and meaning-preserving variations.
  • Progressive curriculum. Multi-stage manifests, per-stage builders, mixed batch sampling with configurable review ratio, and stage-scoped JSONL export.
  • Exercise modules. Built-in catalogs and generators for language basics, math, logic, and coding drills — wired into the curriculum registry.
  • Queryable storage. MemoryDatasetStore for tests; SqliteDatasetStore for large persistent corpora with indexed filters and batched bulk insert.
  • Training streams. Dataset and CurriculumDataset yield batches without loading the full corpus into memory.
  • Reproducibility. Dataset + version, pipeline version, generator versions, and seeds; optional failIfVersionExists guard against silent overwrites.
  • JSONL interchange. DatasetExporter / DatasetImporter preserve metadata, provenance, variation fields, and timestamps.
  • Tested. Unit, integration, and curriculum end-to-end coverage over memory and SQLite stores.

Architecture

           Sources (Memory / Directory / JSONL)
                          │
                    DatasetPipeline
              generate → vary → validate
                          │
                    DatasetStore
                 (Memory / SQLite)
                          │
              Query ──► Dataset / CurriculumDataset
                          │
                    trainer batches

lib/
├── llm_dataset.dart              # public barrel export
└── src/
    ├── model/                    # DatasetEntry, provenance, tool calls
    ├── source/                   # MemorySource, DirectorySource, JsonlSource
    ├── generator/                # DatasetGenerator + built-ins
    ├── variation/                # rule-based + translation variations
    ├── validation/               # composable validators
    ├── pipeline/                 # DatasetPipeline orchestration
    ├── store/                    # Memory + SQLite stores, lifecycle
    ├── query/                    # fluent DatasetQuery filters
    ├── serialization/            # JSON / JSONL codec
    ├── training/                 # Dataset batch streaming
    ├── text_exercise/            # phrases, paragraphs, template store
    ├── curriculum/               # manifest, builder, mixing, lifecycle
    ├── language_basics/          # vocabulary / grammar drills
    ├── math_exercise/            # arithmetic drills
    ├── logic_exercise/           # logic pattern drills
    └── coding_exercise/          # code pattern drills

Getting started

dependencies:
  llm_dataset:
    git:
      url: https://github.com/OmnyGrid/llm_dataset.git

Requires Dart ^3.13. The package uses dart:io and sqlite3 for persistent storage; run on VM targets (not web).

Usage

Pipeline quick start

The snippet at the top builds a memory-backed dataset from one source document, applies rule-based variations, validates, stores, and streams training batches. Swap MemoryDatasetStore for SqliteDatasetStore('corpus.db') for persistence.

Inspect variation groups after the run:

final group = await store.query().limit(1).toList();
final variations = await store
    .query()
    .variationGroup(group.single.variationGroup)
    .toList();
for (final entry in variations) {
  print('${entry.variationIndex}: ${entry.input}');
}

Sources

  • MemorySource — in-process documents
  • DirectorySource — stream files from disk (optional extension filter)
  • JsonlSource — line-by-line source documents (id + content/text)

Generators

Built-ins produce canonical entries and do not depend on an LLM SDK. Supply your own DatasetGenerator to call an external model. See example/custom_generator.dart and example/builtin_generators.dart.

Variations

RuleBasedVariationGenerator applies paraphrase / formal / casual / translation / difficulty / format transforms, preserves variationGroup, assigns indexes, and sets provenance.parentEntryId. Pass seed for deterministic strategy ordering.

For real translation variations, use TranslationVariationGenerator:

variations: [
  CallbackTranslationVariationGenerator(
    targetLanguage: 'es',
    targetLanguages: ['fr', 'de', 'es'], // merged + uniquified → es, fr, de
    translate: (text, {required sourceLanguage, required targetLanguage}) async {
      return await myMtClient.translate(text, from: sourceLanguage, to: targetLanguage);
    },
  ),
],

See example/translation_target_languages.dart for normalizeTargetLanguages usage (exported from the public API).

Validation

Composable validators: empty content, duplicates (stream + store), metadata, language allow-list, length, approximate token count. The pipeline rejects invalid entries and continues; store failures propagate. See example/validation_and_rejections.dart.

Storage & query

final entries = await store
    .query()
    .dataset('general')
    .language('en')
    .variationSelection(VariationSelection.canonicalOnly)
    .sample(100, seed: 1)
    .toList();

SqliteDatasetStore indexes by dataset, version, type, language, variation group, source, created time, and parent. Use configureForBulkInsert() and addAllBatched() when streaming millions of rows.

JSON / JSONL

DatasetExporter / DatasetImporter stream entry JSONL with metadata, provenance, variation fields, and timestamps preserved. See example/jsonl_import_export.dart.

Dataset lifecycle

final lifecycle = DatasetLifecycle(store);
final datasets = await lifecycle.listDatasets();
final versions = await lifecycle.listVersions('geography');
await lifecycle.exportJsonl('out.jsonl', dataset: 'geography', version: 'v1');
await lifecycle.deleteDataset('geography', version: 'v0');

Record dataset + datasetVersion, pipelineVersion, generator versions, and seeds alongside every export. Both DatasetPipeline and CurriculumBuilder refuse to silently overwrite an existing dataset version when failIfVersionExists is true (default). Pass failIfVersionExists: false to rebuild into the same version.

Progressive curriculum training

Build multi-stage datasets for shallow progressive layer training (e.g. language basics → phrases → paragraphs → math → logic → coding). Plateau detection and layer freezing stay in your trainer; this package tags, builds, queries, and exports phase slices.

final manifest = await CurriculumManifest.loadFile('example/curriculum/curriculum.json');
final store = SqliteDatasetStore('curriculum.db');
await CurriculumBuilder(manifest: manifest, store: store).buildAll();

final curriculum = CurriculumDataset(store: store, manifest: manifest);
await for (final batch in curriculum.batchesPhase('phrases', 32, seed: 42)) {
  // trainer.train(batch);
}

await CurriculumLifecycle(store: store, manifest: manifest)
    .exportStageJsonl('phrases', 'phrases.jsonl');

Each entry is tagged with curriculumStage and related metadata. Mixed stages interleave review entries at reviewRatio per batch. Pass variationOptions: CurriculumVariationOptions(variationsPerEntry: 2) to CurriculumBuilder to emit meaning-preserving variations for text sources. See example/curriculum/README.md.

Text exercises & phrase templates

Phrase and paragraph generators, JSON-backed template stores, and full combinatorial expansion across category words, lexicon synonyms, and structural variants. See example/text_exercise.dart and example/phrase_template_categories.dart.

LLM adapters (no SDK in this package)

See example/adapters/README.md for generator and variation adapters that call your model via async callbacks. Local LLM examples: example/local_llm_translation.dart, example/lm_studio_translation.dart.

Examples

Runnable demos live under example/. From the repo root:

dart run example/<file>.dart
Example What it shows
llm_dataset_example.dart Minimal memory pipeline + training stream
end_to_end.dart JSONL source → SQLite → lifecycle → export → batches
jsonl_import_export.dart DatasetExporter / DatasetImporter round-trip
query_and_sampling.dart Filters, canonical-only, one-per-group, sample, shuffle
directory_source.dart DirectorySource with extension filter
tool_call_chat.dart ToolCallGenerator and metadata['messages']
custom_generator.dart Custom DatasetGenerator implementation
builtin_generators.dart All built-in generators on one document
validation_and_rejections.dart CompositeValidator, rejections, version guard
translation_variation.dart TranslationVariationGenerator and callbacks
translation_target_languages.dart Merge + uniquify targetLanguage / targetLanguages
translation_client_providers.dart Custom TranslationClient provider implementations
local_llm_translation.dart Translation via local LLM API (Ollama / LM Studio)
lm_studio_translation.dart LM Studio at 127.0.0.1:1234 — multi-language variation expansion
text_exercise.dart EN/PT phrases & paragraphs with meaning-preserving variations
curriculum_training.dart Self-contained curriculum build, mixed batches, and JSONL export
phrase_template_categories.dart Full combinatorial expansion with per-entry progress logging
curriculum/ Progressive curriculum manifest, build, and mixed-phase training demo
pipeline_store_errors.dart PipelineStoreErrorPolicy.continueProcessing
adapters/ LLM-backed generator and variation adapters

Quick picks:

dart run example/query_and_sampling.dart
dart run example/jsonl_import_export.dart
dart run example/end_to_end.dart
dart run example/curriculum_training.dart

Running the example and tests

dart pub get
dart format lib test example
dart analyze --fatal-infos --fatal-warnings .
dart test

GitHub Actions runs format, analyze, dependency validation, doc dry-run, publish dry-run, and VM tests with coverage uploaded to Codecov. For private repos, add a CODECOV_TOKEN repository secret (from the Codecov project settings).

Author

Graciliano M. Passos: gmpassos@GitHub.

License

Apache License - Version 2.0

Libraries

llm_dataset
Infrastructure for generating, storing, querying, and consuming LLM training datasets.