CoolFace
Datasetpublic

rufatronics/hausa-10phase-synthetic-training-corpus

Hausa 10-Phase Synthetic Training Corpus Summary This repository contains a large, structured, synthetic Hausa-language corpus organized into 10 curriculum phases. The curriculum moves from beginner greetings and everyday services to procedure explanation, intent classification, text transformation, contextual reasoning, domain question answering, structured extraction, evidence-grounded question answering, and safety-oriented robustness tasks. The corpus contains… See the full description on the dataset page: https://huggingface.co/datasets/rufatronics/hausa-10phase-synthetic-training-corpus.

sourceHugging Facecc-by-4.0updated 1mo agoView on Hugging Face
0likes19downloads
Dataset Card

Hausa 10-Phase Synthetic Training Corpus

Summary

This repository contains a large, structured, synthetic Hausa-language corpus organized into 10 curriculum phases. The curriculum moves from beginner greetings and everyday services to procedure explanation, intent classification, text transformation, contextual reasoning, domain question answering, structured extraction, evidence-grounded question answering, and safety-oriented robustness tasks.

The corpus contains 75,693,486 records and 100,000,063,399 validated raw UTF-8-equivalent bytes, distributed across 10 logical batches. Each logical batch contains ten approximately 1 GB raw phase shards, giving a total of 100 phase shards. The raw logical size and the compressed Hugging Face storage size are reported separately because gzip compression makes them substantially different.

Important: This is a synthetic, template/combinatorial corpus. It is not a collection of naturally occurring Hausa text, and its size must not be interpreted as 100 GB of unique linguistic knowledge.

Intended uses

The corpus is intended for Hausa NLP prototyping, data-pipeline testing, curriculum-learning experiments, structured prediction, instruction-format experiments, safety-task prototyping, and baseline model development. It can also be used to test loading, sharding, validation, deduplication, and evaluation infrastructure for a low-resource language.

The corpus should not be used by itself as evidence of fluent Hausa generation, as a substitute for community-reviewed language resources, or as the sole training source for a production model. High-stakes medical, legal, financial, civic, or safety decisions require independently verified sources and qualified human review.

Data generation methodology

The large corpus was produced by the deterministic streaming generator in src/stream_10gb_batch.py. The generator combines phase-specific Hausa phrase banks with controlled values for names, places, days, times, topics, record numbers, and sample codes. It writes one JSON record at a time to gzip-compressed JSONL, so the full corpus does not need to be held in memory or stored locally at once.

Each record receives a batch-specific and phase-specific identifier. The generator records the phase, task family, domain, difficulty, input, expected output, labels, context, safety profile, split, provenance fields, and quality-check metadata. Each phase was generated until its raw UTF-8-equivalent payload reached approximately 1,000,000,000 bytes. The resulting shards were validated, uploaded, checked remotely, and then removed from local staging storage.

The initial small pilot included model-assisted expansion experiments. The 100 GB production corpus was generated using streamed_combinatorial_template_v1. This distinction matters when interpreting diversity and linguistic quality.

Curriculum phases

PhaseDifficultyPrimary task coverage
1BeginnerGreetings, simple questions, everyday replies, and basic dialogue
2ElementaryShopping, transport, directions, prices, time, and appointments
3Lower intermediateProcedures, explanations, household topics, agriculture, and cautious advice
4IntermediateIntent, sentiment, emotion, urgency, topic classification, and hard negatives
5Upper intermediateParaphrase, summarization, simplification, question generation, and normalization
6AdvancedContextual reasoning, temporal references, entities, relationships, and contradictions
7Domain advancedAgriculture, education, commerce, technology, and public-service language
8Structured advancedEntity, location, date, event, relation, and JSON-style structured extraction
9ExpertEvidence-grounded question answering and separation of stated facts from assumptions
10Research frontierRobustness, safe refusal, uncertainty handling, evaluation, and adversarial task framing

Record schema

Every production record follows the same broad JSON structure:

FieldDescription
record_idDeterministic batch- and phase-specific identifier
phase_idCurriculum phase from 1 through 10
task_familyTask type such as beginner_dialogue, domain_qa, or structured_extraction
domainTopic or task domain
languageha
scriptLatin orthography in this production release
difficultyCurriculum difficulty label
instructionTask instruction in Hausa
inputSynthetic Hausa prompt or context
outputExpected response or structured output
labelsTask-specific labels and metadata
contextStructured values used for the example
safety_profileSafety category assigned by the generator
syntheticAlways true for production records
generation_methodGeneration method identifier
splittrain, validation, or test
quality_checksAutomated validation flags

Measured metrics

The following metrics were computed from the 100 production validation manifests and a reproducible random sample of three records from each phase in Batch 010.

MetricResultInterpretation
Logical batches10Ten retained 10 GB logical batches
Curriculum phases10All phases represented in every batch
Records75,693,486Validated JSON records
Raw logical payload100,000,063,399 bytesApproximately 100.000 GB decimal
Raw logical payload93.132 GiBBinary-unit equivalent
Production validation manifests100One per batch and phase
Validation errors0No automated schema or accounting errors reported
Reported unique record IDs75,693,486IDs unique within the validated shard reports
Reported unique text hashes75,693,486Exact text-hash uniqueness within the validated shard reports
Train split68,124,536 records, 90.0005%Main training split
Validation split3,784,493 records, 4.9998%Development split
Test split3,784,457 records, 4.9997%Evaluation split
Compressed phase-shard storage3,785,206,316 bytesApproximately 3.785 GB decimal
Compression factor26.42×Raw logical bytes divided by compressed phase-shard bytes
Random sample size30 recordsThree records from each phase
Sample task families10All ten curriculum task families observed
Sample unique domains23Domain variation in the 30-record sample
Sample missing required fields0Required schema fields present in all sampled records
Phase-8 valid JSON outputs3 / 3All sampled structured outputs parsed as JSON objects
Sample unique input hashes30 / 30No exact input duplicates in the sample
Sample unique output hashes26 / 30Repeated response patterns are present

The validation metrics establish technical integrity, not linguistic correctness. Exact hash uniqueness should not be confused with semantic diversity: the generator embeds a unique sample code in each prompt, so near-duplicate templates can still exist. Independent Hausa-speaker review, semantic deduplication, and naturalness scoring remain necessary for a production-quality corpus.

Quality assessment and limitations

The corpus has strong technical properties: a stable schema, deterministic identifiers, reproducible batch generation, phase coverage, gzip integrity, split accounting, and machine-readable manifests. It is therefore useful for testing training infrastructure and controlled task-format experiments.

Its principal weakness is linguistic and semantic diversity. Many records are produced by recombining a finite set of templates. A changed name, location, time, topic, or record code can make a record technically different while leaving the underlying sentence pattern nearly unchanged. Some combinations may sound unnatural to a Hausa speaker, and some task labels are used as synthetic topic values. The metadata fields language_review and safety_review remain pending until qualified human reviewers assess the records.

The production release uses general Hausa Latin orthography. It should not be assumed to represent all regional varieties, dialects, registers, or Ajami writing. It also contains no claim that synthetic outputs are factual statements about the people, places, dates, or topics appearing in the records.

Recommended use in training

For responsible experimentation, use the corpus as one component of a broader mixture. Begin with a small sample and inspect language quality. Deduplicate semantically across batches, not only by exact string hash. Keep validation and test examples isolated from template families used in training. Weight the corpus alongside human-authored and community-reviewed Hausa text, and evaluate the resulting model with native-speaker judgments covering fluency, adequacy, factuality, dialect fit, harmfulness, and instruction-following.

Loading the data

The repository is available at:

<https://huggingface.co/datasets/rufatronics/hausa-10phase-synthetic-training-corpus>

The phase shards are stored beneath data/batch_001 through data/batch_010. The validation manifests are stored beneath the corresponding batch directories. Because this release is intentionally sharded and compressed, users may choose a subset of batches or phases for experiments rather than loading the full corpus at once.

License and attribution

The repository metadata declares cc-by-4.0 for the synthetic artifacts. Users should verify that license suitability for their intended deployment and should retain the dataset card and provenance information when redistributing derived artifacts.

Collection

This dataset is included in the public VelkroLM Hausa collection:

<https://huggingface.co/collections/VelkroLM/hausa>

Citation

bibtex
@dataset{hausa_10phase_synthetic_training_corpus,
  title = {Hausa 10-Phase Synthetic Training Corpus},
  author = {Manus AI},
  year = {2026},
  publisher = {Hugging Face},
  repository = {rufatronics/hausa-10phase-synthetic-training-corpus},
  language = {Hausa},
  note = {Synthetic streamed combinatorial corpus with ten curriculum phases}
}