job_type_index.md

EasyDistill 2 Job Type Index

This page maps every supported CLI job_type to a representative config file and the relevant documentation. Config paths are shown for the pai_token backend; each file has a matching _pai_eas variant unless noted otherwise.

For an interactive list, run easydistill --list-jobs.

Basic distillation

job_type	Purpose	Representative config	Documentation
instruct_distill	Generate teacher responses for seed instructions and build SFT data.	configs/basic/instruct_distill_pai_token.yaml	instruction_distillation.md
cot_distill	Generate chain-of-thought reasoning traces and build SFT data.	configs/basic/cot_distill_pai_token.yaml	cot_distillation.md
mm_instruct_distill	Generate teacher responses for (image, instruction) pairs.	configs/basic/mm_instruct_distill_pai_token.yaml	mm_distillation.md
mm_cot_distill	Generate visual chain-of-thought reasoning traces.	configs/basic/mm_cot_distill_pai_token.yaml	mm_cot_distillation.md
t2v_distill	Generate videos from seed prompts (T2V/I2V) and build SFT data.	configs/basic/t2v_distill_pai_token.yaml	t2v_distillation.md

End-to-end pipelines

job_type	Purpose	Representative config	Documentation
advanced_instruct_distill	Expand → generate → judge → filter → SFT.	configs/pipeline/advanced_instruct_distill_pai_token.yaml	advanced_instruct_distill.md
balanced_instruct_distill	Balance categories, generate, evaluate, filter, build SFT.	configs/pipeline/balanced_instruct_distill_pai_token.yaml	balanced_instruct_distill.md
augmented_instruct_distill	Refine seeds, generate multiple responses, evaluate, filter, build SFT.	configs/pipeline/augmented_instruct_distill_pai_token.yaml	augmented_instruct_distill.md
advanced_cot_distill	Generate CoT, score with RV/CD, mix by difficulty, build SFT.	configs/pipeline/advanced_cot_distill_pai_token.yaml	cot_rvcd_mixer.md
advanced_mm_distill	Multi-modal generate → eval → filter → SFT.	configs/pipeline/advanced_mm_distill_pai_token.yaml	mm_distillation.md
advanced_mm_cot_distill	Visual CoT with RV/CD/correctness scoring.	configs/pipeline/advanced_mm_cot_distill_pai_token.yaml	mm_cot_distillation.md
advanced_mm_cot_distill (OmniThoughtV)	Same job_type; reproduces the OmniThoughtV recipe.	configs/pipeline/omnithoughtv_mm_cot_distill_pai_token.yaml	mm_cot_distillation.md
advanced_t2i_distill	Prompt optimization → T2I generation → VLM judge → filter → SFT.	configs/t2i/advanced_t2i_distill_wanx.yaml	t2i_distillation.md
advanced_t2v_distill	Prompt optimization → video generation → video eval → filter → SFT.	configs/pipeline/advanced_t2v_distill_pai_token.yaml	t2v_distillation.md
pe_rewrite_distill	Plan/rewrite/reflection → judge → filter → SFT for prompt rewriting.	configs/pipeline/pe_rewrite_distill_from_seeds_pai_token.yaml	pe_rewrite.md
agent_distill	Synthesize tool-use tasks and build agent trajectory SFT/DPO data.	configs/pipeline/agent_distill_pai_token.yaml (SFT) or configs/pipeline/agent_distill_dpo_pai_token.yaml (DPO)	agent_distillation.md
search_agent_distill	Evolve seed QA into multi-hop search tasks and build SFT data.	configs/pipeline/search_agent_distill_pai_token.yaml	search_agent_distillation.md

Preference data

job_type	Purpose	Representative config	Documentation
dpo_data_build	Generate candidates, score them, and build DPO preference pairs.	configs/preference/dpo_instruct_pai_token.yaml (set dpo.task_type: instruct) or configs/preference/dpo_cot_pai_token.yaml (set dpo.task_type: cot)	dpo_distillation.md

Text rewrite and synthesis operators

job_type	Purpose	Representative config	Documentation
instruction_expansion	Synthesize new instructions from seed examples.	configs/rewrite/instruction_expansion_pai_token.yaml	instruction_balancing.md
seed_anchored_expansion	Expand each seed into same-scenario instructions with dedup and lineage.	configs/rewrite/seed_anchored_expansion_pai_token.yaml	instruction_balancing.md
instruction_refinement	Rewrite and improve existing instructions.	configs/rewrite/instruction_refinement_pai_token.yaml	instruction_balancing.md
instruction_response_extraction	Extract instruction/response pairs from raw text.	configs/rewrite/instruction_response_extraction_pai_token.yaml	instruction_balancing.md
agentic_rewrite	Rewrite prompts via a plan → rewrite → reflection teacher agent chain.	configs/rewrite/agentic_rewrite_pai_token.yaml	pe_rewrite.md
cot_long2short	Simplify existing CoT reasoning traces.	configs/rewrite/cot_long2short_pai_token.yaml	cot_distillation.md
cot_short2long	Extend existing CoT reasoning traces with more detail.	configs/rewrite/cot_short2long_pai_token.yaml	cot_distillation.md
mm_cot_long2short	Simplify multi-modal CoT reasoning traces.	configs/rewrite/mm_cot_long2short_pai_token.yaml	mm_cot_distillation.md
mm_cot_short2long	Extend multi-modal CoT reasoning traces with more detail.	configs/rewrite/mm_cot_short2long_pai_token.yaml	mm_cot_distillation.md

PE rewrite pipeline stages

job_type	Purpose	Representative config	Documentation
pe_rewrite_eval	Score prompt rewrites with a multi-dimension LLM judge.	configs/rewrite/pe_rewrite_eval_pai_token.yaml	pe_rewrite.md
pe_rewrite_filter	Filter judged rewrites by score thresholds and top ratio.	configs/rewrite/pe_rewrite_filter.yaml	pe_rewrite.md
pe_rewrite_build_sft	Build SFT samples from filtered rewrites.	configs/rewrite/pe_rewrite_build_sft.yaml	pe_rewrite.md

Evaluation operators

job_type	Purpose	Representative config	Documentation
instruct_eval	Run LLM-as-judge evaluation on instruction/response pairs.	configs/eval/instruct_eval_pai_token.yaml	data_formats.md
cot_eval	Run LLM-as-judge evaluation on CoT reasoning traces.	configs/eval/cot_eval_pai_token.yaml	data_formats.md
mm_instruct_eval	Run LLM-as-judge evaluation on multi-modal instruction responses.	configs/eval/mm_instruct_eval_pai_token.yaml	data_formats.md
mm_cot_eval	Run LLM-as-judge evaluation on multi-modal CoT traces.	configs/eval/mm_cot_eval_pai_token.yaml	data_formats.md
t2i_eval	Run VLM-as-judge evaluation on generated images.	configs/eval/t2i_eval_pai_token.yaml	data_formats.md
t2i_single_model_eval	Single-teacher T2I evaluation with a dimension pool judge.	configs/eval/t2i_ti2i/t2i_single_model_pai_token.yaml	t2i_ti2i_eval.md
t2i_multi_model_eval	Multi-teacher T2I evaluation with cross-model debate.	configs/eval/t2i_ti2i/t2i_multi_model_pai_token.yaml	t2i_ti2i_eval.md
ti2i_single_model_eval	Single-teacher TI2I evaluation with a dimension pool judge.	configs/eval/t2i_ti2i/ti2i_single_model_pai_token.yaml	t2i_ti2i_eval.md
ti2i_multi_model_eval	Multi-teacher TI2I evaluation with cross-model debate.	configs/eval/t2i_ti2i/ti2i_multi_model_pai_token.yaml	t2i_ti2i_eval.md
t2v_eval	Run T2V video evaluation (precheck, VLM judge, optional omni check).	Reuses a T2V pipeline config with resume/eval enabled, or configs/eval/t2v/vlm_dimensions.yaml for dimension definitions.	t2v_distillation.md

T2I/T2V generation operators

job_type	Purpose	Representative config	Documentation
prompt_optimize	Optimize seed T2I prompts into rich, descriptive prompts.	configs/t2i/prompt_optimize_pai_token.yaml	t2i_distillation.md
t2i_generation	Generate images from prompts via a T2I backend (no SFT building).	configs/t2i/t2i_generation_wanx.yaml	t2i_distillation.md
t2v_prompt_optimize	Two-stage T2V/I2V prompt optimization.	Reuses configs/basic/t2v_distill_pai_token.yaml with prompt-optimization stages enabled.	t2v_distillation.md
t2v_generation	Generate videos from prompts via a T2V backend (no SFT building).	Reuses configs/basic/t2v_distill_pai_token.yaml with generation stage enabled.	t2v_distillation.md

Notes

• Config paths ending in _pai_token.yaml have _pai_eas.yaml counterparts under the same directory.
• T2I configs use backend-specific variants: _wanx.yaml, _qwen_image.yaml, or _pai_diffusion.yaml.
• Some standalone operators do not have dedicated config files; they reuse pipeline configs and are enabled by setting the appropriate stage flags or by resuming from an intermediate JSONL.