v1.0.3

[AronDaron]

v1.0.3 - Recovery

Fixed

• Failed jobs are now resumable from /history. Previously only interrupted and cancelled jobs could be resumed — a failed job (e.g. transient OpenRouter glitch returning 200 OK with malformed body, no choices) wiped 6+ hours of work with no way to continue. The Resume button now appears for failed status too, picks up progress from the examples table, and continues from where it stopped. Works retroactively for any historical failed jobs.
• NotificationsPopover badge now pulses for both interrupted and failed (any unexpected interruption), not just interrupted. User-cancelled jobs continue to live silently in /history.
• Strict turn-count validation for sharegpt/chatml jobs. The schema validator previously only checked len(turns) >= 2, so a job configured for conversation_turns=2 could silently accept 1-turn or 3-turn dialogues from the model. Discovered after a 510-example sharegpt run had 9 off-spec examples (1.8%) — 8 with 3 turns, 1 with 1 turn. The validator now rejects mismatched length with reason turn_count_mismatch, triggering the same retry/skip mechanism as other validation failures. Existing rows in DB are not re-validated — merge and export only strip keys, never re-run the schema check, so historical datasets stay intact. Alpaca format is unaffected (no turn list).

Changed

• /api/jobs/resumable and /api/jobs/{id}/resume endpoints accept failed status (was interrupted/cancelled only).
• Frontend resumable badge logic computes unexpectedCount locally from the jobs list (interrupted + failed) instead of relying on the legacy interrupted_count field. The API field is preserved for backward compatibility but no longer drives UI state.
• validate_example() and is_valid_example() accept an optional expected_turns: int | None parameter; default None keeps the legacy loose check (>= 2 floor only). The job runner forwards config.conversation_turns to it on every generation attempt.

---

This discussion was created from the release v1.0.3.