This public-gated dataset repository stores KinyaLM data artifacts for team review. Visitors can see the dataset page, but file access requires a logged-in Hugging Face account and gated access acceptance.

Current draft batches (1,000 rows total):

sft-drafts-2026-07-13-batch-001: 286 draft rows
sft-drafts-2026-07-13-batch-002: 714 draft rows

Each batch includes:

draft SFT JSONL,
a master review TSV and any review shards,
a review package zip,
manifests with checksums and review status.

Synthetic distillation review batch:

sft-distillation-production-1000-v3-final: 1,000 deterministic-gate-passing candidates, 863 critic-pass recommendations, and 137 critic-repair recommendations.

The distillation batch includes complete teacher and critic provenance, the provider-neutral job manifests, contamination-gated drafts, a quality summary, and separate review queues for TessyMugisha and BonheurByiringiro. Its 86.3% critic acceptance rate is below the 90% scale threshold, so the next 3,000-row expansion is paused until fluent reviewers resolve the repair queue.

Status

This is not an approved training dataset. Batch rows are draft examples marked split=draft and review_status=needs-review.

Review Rule

Rows can train only after fluent-speaker review promotes them to:

review_status=approved
split=train or split=validation

Use the project promotion script after review:

python3 scripts/promote_reviewed_sft.py \
  --draft-jsonl ~/KinyaLMData/drafts/sft-drafts-2026-07-13-batch-001.jsonl \
  --review-tsv ~/KinyaLMData/reviewed/sft-drafts-2026-07-13-batch-001.review.tsv \
  --out-dir ~/KinyaLMData/approved/sft-drafts-2026-07-13-batch-001

Source And License Notes

These batches are synthetic draft data generated for review. They are not redistributed as approved public training data until the team confirms quality and release terms.

Downloads last month: 126

Total file size:

114 MB