[yaoxiao67]

This PR adds training-data preparation operators and opt-in resilience for local JSONL tasks. Invalid content can be inspected before explicit filtering, committed pipeline outputs can be resumed, and failed independent model requests can be replayed.

Changes

• Add format-specific QA, Alpaca and text ShareGPT messages builders with fixed field signatures and source-field preservation.
• Separate source/conversation/preference evaluation from explicit filtering. Conversion preserves row count and raises on invalid content.
• Keep preference field mapping as a Python utility, preserving both answers.
• Add opt-in malformed-JSONL isolation, validated batch/stream checkpoints and recorded HTTP request budgets/retries.
• Provide an independent recorded-generation runner with aligned success/failure outcomes and an explicit successful-results export.
• Include bilingual operator descriptions, migration notes and three CPU examples.

Compatibility and classification

• Existing upstream defaults and serving return types are preserved; no runtime dependencies were added.
• The unpublished core_text interfaces from the first PR revision have been removed and their examples/tests migrated.
• Please review the proposed domain generate placement for structured messages builders and the text_preference domain for preference evaluation/filtering.
• Enhanced recovery targets local JSONL and a single writer. Deduplication, distributed storage, multimodal data and conversational DPO are outside this change.

Validation

• macOS arm64 / Python 3.11.17: python -m pytest -m 'cpu or api' -q -rs — 425 passed, 48 deselected.
• Public-data, tool/error, SIGKILL recovery and localhost HTTP/replay tasks passed from source and an independently installed wheel.
• 56,002 public records: 55,973 accepted and 29 empty-content errors; source/wheel training exports matched, with content unchanged from the preceding revision.
• Installed dependency checks and sdist/wheel source checks passed. HTTP faults used a local server; no real model API was called.

Please review the public interfaces, taxonomy and opt-in recovery boundaries. Native platform matrix results remain for upstream CI.

[yaoxiao67]

I have revised the implementation against the framework conventions:

• Moved training-format rules out of core_text into SFT, conversation and preference domains.
• Split format-specific operators with fixed field signatures so compile() can detect missing default and enabled optional inputs.
• Separated evaluation, explicit filtering and conversion; converters no longer drop invalid records implicitly.
• Added bilingual descriptions, parameter defaults, error semantics and migration examples.
• Replaced RecordedPromptedGenerator with an explicit task runner, keeping aligned failure outcomes separate from the successful-results export.

Local CPU/API selection: 425 passed. Public-data, tool/error, real process-interruption and localhost HTTP/replay tasks passed from both source and an installed wheel.

Please review two classification proposals: domain generate for structured messages builders that preserve source columns, and text_preference for preference evaluation/filtering. These choices are proposals for review and can be adjusted to the project's preferred taxonomy.