Recipes

The same listing llmclick recipes and GET /api/system/recipes return. The last column names the values a config may write for keys that resolve through a registry: architectures, row sources, rewards, decision heads, teachers, generators and evolution operators.

Recipe	Stages	Registry keys the config may use
data_speech	seeds, prompts, synthesize, verify, select	teacher: anthropic, local, ollama, openai; generator: document_qa, labelled_requests, passthrough, persona_task, topic_questions, utterances
data_synthetic	seeds, prompts, evolve, respond, verify, select	teacher: anthropic, local, ollama, openai; generator: document_qa, labelled_requests, passthrough, persona_task, topic_questions, utterances; evolver: breadth, code_complexity, code_misdirection, code_requirement, complicate, concretize, constraints, deepen, easier, reasoning
embedding_contrastive	data, train, validate	architecture: bi_encoder; source: delimited, huggingface, jevbench, local_jsonl; method_keys: query_instruction, temperature
evaluation_benchmark	benchmark, report	checkpoint: validate, train, base, <directory>; preset: korean, small_general
evaluation_compare	rows, score, benchmark, decision, mteb, report	checkpoint: validate, train, base, <directory>; preset: korean, small_general; source: delimited, huggingface, jevbench, local_jsonl
evaluation_custom	rows, score, report	checkpoint: validate, train, base, <directory>; source: delimited, huggingface, jevbench, local_jsonl
evaluation_decision	rows, decision, report	checkpoint: validate, train, base, <directory>; source: delimited, huggingface, jevbench, local_jsonl; think: off, on, both
evaluation_embedding	mteb, report	checkpoint: validate, train, base, <directory>
evaluation_wakeword	clips, traces, report	checkpoint: validate, train, <directory>
llm_decision_cispo	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys: anchor_weight, calibration, clip, eval_think, group_size, head_lr, iterations, length_hinge, length_penalty_cap, max_think, rollout_weight, temperature; head: pointer, readout
llm_decision_sft	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys: calibration, eval_think, head_lr, max_think, think_fraction; head: pointer, readout
llm_dpo	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys: beta
llm_grpo	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys: beta, clip, group_size, iterations, max_new_tokens, reward, temperature; reward: contains, exact_match, last_number
llm_instruction	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys: system
llm_sft	data, train, validate	architecture: hybrid, transformer; source: delimited, huggingface, jevbench, local_jsonl; method_keys:
tts_piper	data, train, validate	quality: high, medium; source: delimited, huggingface, jevbench, local_jsonl
wakeword	data, features, train, validate	source: delimited, huggingface, jevbench, local_jsonl