data_juicer.ops.mapper.dialog_quality_llm_base module

Internal base classes for dialog/agent turn- and trace-quality LLM mappers.

Each concrete mapper performs one API call per sample and writes meta[META_KEY] with score (1–5), reason, and eval_kind. Inputs are composable: set history_key / query_key / response_key or text_key to match your dataset. After agent_dialog_normalize_mapper, dialog_history[-1] matches query/response; build_dialog_turn_eval_user_content deduplicates so that final turn is not repeated under both “Earlier turns” and “Current” (which previously biased non-repetition and topic-shift judges). Override rubrics by subclassing or forking the mapper; rubric text stays English-friendly. Pass preferred_output_lang="zh" (YAML) so JSON reason and instructions use Chinese; "en" keeps English (default).