Repository navigation
💫 Release v0.10.5
For thumbnail navigation fixes, use the follow-up 0.10.6 release.
OpenAI chat exports to LLaMA-Factory now preserve messages/role/content and generate matching columns and role tags. Redirected dt export output is one JSON summary; success messages and file paths go to stderr and honor --quiet.
- Restore streaming CSV/TSV reads with current Polars using
collect_batches. - Fix the ShareGPT generator so all 100 demonstration rows convert and validate; keep malformed data in an explicit error fixture.
- Clarify dry-run limits and align the training-data workbench positioning.
- Add reproducible synthetic workflows, two short terminal demos, launch drafts and daily metrics tooling in the growth execution pack.
Validation: 1307 tests passed; changed-file Ruff/Black and wheel/sdist checks passed; independent QA passed. LLaMA-Factory 0.9.5's real parser verified standard text-chat mappings. An isolated wheel installation ran the complete 308-to-252-sample workflow successfully. Existing whole-repository hatch run lint:all failures remain; GPU, tool and multimodal training were not validated.