Config schema

A config is one YAML with a pipeline block (recipe, name, seed, output_dir, device, stages) and one section per concern. Every section rejects keys it does not know, so a typo is an error, not a silently ignored setting. The tables below come straight from the Pydantic classes; nested keys are written with dots.

data_speech

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
teacher	Keyed | null		The model generators ask; null when the generator asks none (passthrough)
teacher.name	string	required
seeds	list[SeedSpec]	required
seeds.kind	string	required	What the seeds are: topic, persona, document, instruction, ...; generators read it
seeds.values	list[string] | null		Inline seed texts
seeds.source	Keyed | null		A row source (local_jsonl, huggingface, ...)
seeds.source.name	string	required
seeds.field	string | null		With source: the column that holds the seed text
seeds.limit	integer | null
prompts	PromptSettings	required
prompts.generator	string	required	Key of GENERATORS: passthrough, topic_questions, persona_task, document_qa
prompts.per_seed	integer	1	Prompts to make from each seed
prompts.params	object (extra keys allowed)		Generator parameters, e.g. skills for persona_task
prompts.constraints	list[string]		Verifiable constraints; one is appended to every prompt
prompts.max_tokens	integer	512
prompts.temperature	number	1.0
voice	VoiceSettings	required
voice.model	HubFile	required	A Piper voice (.onnx); its .onnx.json is taken from beside it
voice.model.name	string	required	Hub repository ID, or a local file or directory
voice.model.revision	string | null		40-character commit; with file, for a Hub repository
voice.model.file	string | null		Path of the file inside the Hub repository
voice.samples	integer	1	Recordings per text; Piper's noise makes each one different
voice.speakers	list[integer] | integer | null		Speakers of a multi-speaker voice, by id, or n for ids 0 to n-1; every text is read by each of them
voice.length_scale	number | null		Above 1 speaks slower; null: the voice's own
voice.noise_scale	number | null		null: the voice's own
voice.noise_w_scale	number | null		null: the voice's own
verify	SpeechVerifySettings
verify.min_seconds	number	0.5
verify.max_seconds	number	15.0	Piper trains on short clips; Whisper hears 30 s at most
verify.max_clipped	number	0.001	Share of samples at full scale
verify.max_silent	number	0.5	Share of silent 20 ms frames
verify.asr	AsrConfig | null		null skips the listening check
verify.asr.name	string	openai/whisper-base	Whisper model on the Hub
verify.asr.revision	string	e37978b90ca9030d5170a5c07aadb050351a65bb	40-character commit
verify.asr.language	string | null		Language code (en, ko, ...); null lets Whisper detect it
verify.max_wer	number	0.2	Word error rate of the transcript against the text
select	SpeechSelectSettings
select.per_text	integer | null	1	Recordings kept per text, best heard first; null: all, for a wake word read by many speakers
select.max_rows	integer | null
licence	string	required	Terms of the voice, and so of every recording it makes; kept in the manifest

data_synthetic

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
teacher	Keyed	required
teacher.name	string	required
seeds	list[SeedSpec]	required
seeds.kind	string	required	What the seeds are: topic, persona, document, instruction, ...; generators read it
seeds.values	list[string] | null		Inline seed texts
seeds.source	Keyed | null		A row source (local_jsonl, huggingface, ...)
seeds.source.name	string	required
seeds.field	string | null		With source: the column that holds the seed text
seeds.limit	integer | null
prompts	PromptSettings	required
prompts.generator	string	required	Key of GENERATORS: passthrough, topic_questions, persona_task, document_qa
prompts.per_seed	integer	1	Prompts to make from each seed
prompts.params	object (extra keys allowed)		Generator parameters, e.g. skills for persona_task
prompts.constraints	list[string]		Verifiable constraints; one is appended to every prompt
prompts.max_tokens	integer	512
prompts.temperature	number	1.0
evolve	EvolveSettings | null
evolve.rounds	integer	1
evolve.operators	dict[str, number]		EVOLVERS keys and their draw weights
evolve.max_tokens	integer	512
evolve.temperature	number	1.0
evolve.judge_gain	boolean	False	Ask the teacher whether the evolved prompt adds information (Evol-Instruct rule 1)
respond	RespondSettings
respond.answers_per_question	integer	1
respond.system	string | null
respond.max_tokens	integer	1024
respond.temperature	number	0.7
verify	VerifySettings
verify.schema_recipe	string | null		A modeling recipe whose check every assembled row must pass
verify.reward	Keyed | null		A REWARDS function scoring the answer against the prompt's answer
verify.reward.name	string	required
verify.majority	last_number | exact | null		Keep answers that agree with the majority of a prompt's answers
verify.judge	JudgeSettings | null
verify.judge.teacher	Keyed | null		Another model family than the generator; null: the generator
verify.judge.teacher.name	string	required
verify.judge.rubric	string	Rate how helpful, correct and complete the answer is for the request.
verify.judge.min_score	number	6
verify.judge.max_tokens	integer	64
verify.sorry_words	integer	80	An answer with 'sorry' and fewer words than this is a failed prompt (Evol-Instruct rule 2)
select	SelectSettings
select.assemble	sft | dpo | grpo | decision	sft
select.decision	DecisionQuestion | null		assemble: decision, or verify.schema_recipe: llm_decision_sft: the question every row asks
select.decision.name	string	route	The question's key in the row's questions
select.decision.instructions	string	required	What the decision model is asked about the state
select.decision.criteria	dict[str, string]	required	Option -> what it means; the labels are these options
select.dedup	DedupSettings
select.dedup.near	number | null	0.75	Share of word n-grams shared with a kept row that makes a duplicate; null: exact only
select.dedup.ngram	integer	5
select.leak	LeakSettings | null
select.leak.against	list[Keyed]	required	Row sources of the evaluation items the rows must not contain
select.leak.against.name	string	required
select.leak.threshold	number	0.5
select.leak.ngram	integer	8
select.margin	number	1.0	dpo: the judge-score gap a chosen and rejected answer must have
select.max_rows	integer | null
budget	BudgetSettings
budget.max_usd	number | null		Stop before a teacher call would exceed this spend
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]

embedding_contrastive

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	BackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.temperature	number	0.05
method.query_instruction	string | null		Task description prepended to every query

evaluation_benchmark

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
decoding	Decoding
decoding.temperature	number	0.0	0: greedy
decoding.max_new_tokens	integer	1024
chat_template	boolean	False	Wrap prompts in the tokenizer's chat template (instruction-tuned models)
thinking	boolean | null		With chat_template: pass enable_thinking to the template (Qwen3, Qwen3.5); true also splits answers at . null: the template's default
model	SourceModel	required
model.experiment	string | null		Experiment directory (output/-) with its config.yaml
model.checkpoint	string	validate	validate (passed validation), train (as trained), base (the weights the experiment started from), or a checkpoint directory
model.name	string | null		Instead of an experiment: a Hugging Face model ID or a local directory
model.revision	string | null		40-character commit of name; required unless it is a directory
model.architecture	string | null		With name: the model catalog entry (e.g. bi_encoder) for recipes that load the model themselves
benchmarks	Benchmarks	required
benchmarks.preset	string | null		One of ['korean', 'small_general']; tasks are added to it
benchmarks.tasks	list[Task]
benchmarks.tasks.name	string	required	An lm-evaluation-harness task or group name
benchmarks.tasks.num_fewshot	integer | null		null: the task's own default
benchmarks.tasks.limit	integer | null		Items per task; null: benchmarks.limit, then all
benchmarks.tasks.chat_template	boolean | null		Overrides chat_template for this task (multiple choice scored by likelihood often reads better without it); null: the config's
benchmarks.limit	integer | null		Items per task unless the task says otherwise; null: all (full runs take hours)

evaluation_compare

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
decoding	Decoding
decoding.temperature	number	0.0	0: greedy
decoding.max_new_tokens	integer	1024
chat_template	boolean	False	Wrap prompts in the tokenizer's chat template (instruction-tuned models)
thinking	boolean | null		With chat_template: pass enable_thinking to the template (Qwen3, Qwen3.5); true also splits answers at . null: the template's default
runs	list[Run]	required	The checkpoints to compare; one of them checkpoint: base
runs.experiment	string | null		Experiment directory (output/-) with its config.yaml
runs.checkpoint	string	validate	validate (passed validation), train (as trained), base (the weights the experiment started from), or a checkpoint directory
runs.name	string | null		Instead of an experiment: a Hugging Face model ID or a local directory
runs.revision	string | null		40-character commit of name; required unless it is a directory
runs.architecture	string | null		With name: the model catalog entry (e.g. bi_encoder) for recipes that load the model themselves
runs.label	string	required
benchmarks	Benchmarks | null		Public benchmarks every run is scored on
benchmarks.preset	string | null		One of ['korean', 'small_general']; tasks are added to it
benchmarks.tasks	list[Task]
benchmarks.tasks.name	string	required	An lm-evaluation-harness task or group name
benchmarks.tasks.num_fewshot	integer | null		null: the task's own default
benchmarks.tasks.limit	integer | null		Items per task; null: benchmarks.limit, then all
benchmarks.tasks.chat_template	boolean | null		Overrides chat_template for this task (multiple choice scored by likelihood often reads better without it); null: the config's
benchmarks.limit	integer | null		Items per task unless the task says otherwise; null: all (full runs take hours)
rows	EvaluationData | null		Rows every run is scored on with its recipe's metrics
rows.sources	list[Keyed]	required	Rows in the format of the experiment's recipe
rows.sources.name	string	required
decision	DecisionRows | null		Decision rows (or JevBench) every run answers, with think settings
decision.think	off | on | both	both	Measure without reasoning (off), with it (on), or both
decision.max_think	integer | null		Reasoning tokens per question; null: the experiment's method.max_think
decision.data	EvaluationData	required
decision.data.sources	list[Keyed]	required	Rows in the format of the experiment's recipe
decision.data.sources.name	string	required
embedding	EmbeddingSettings | null		MTEB tasks every run is scored on
embedding.tasks	list[string]	required	MTEB task names, e.g. STSBenchmark, NFCorpus, Banking77Classification
embedding.max_length	integer | null		Tokens per text; null: the experiment's training.max_length
axes	dict[str, list[string]]		Chart axis -> entries averaged, each 0..1: a benchmark name, rows:<metric>, decision:<mode>:<metric> or embedding:<task>

evaluation_custom

evaluation_custom: one model, any rows, the metrics of the recipe that trained it.

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
model	SourceModel	required
model.experiment	string | null		Experiment directory (output/-) with its config.yaml
model.checkpoint	string	validate	validate (passed validation), train (as trained), base (the weights the experiment started from), or a checkpoint directory
model.name	string | null		Instead of an experiment: a Hugging Face model ID or a local directory
model.revision	string | null		40-character commit of name; required unless it is a directory
model.architecture	string | null		With name: the model catalog entry (e.g. bi_encoder) for recipes that load the model themselves
data	EvaluationData	required
data.sources	list[Keyed]	required	Rows in the format of the experiment's recipe
data.sources.name	string	required

evaluation_decision

Key	Type	Default	Description
think	off | on | both	both	Measure without reasoning (off), with it (on), or both
max_think	integer | null		Reasoning tokens per question; null: the experiment's method.max_think
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
model	SourceModel	required
model.experiment	string | null		Experiment directory (output/-) with its config.yaml
model.checkpoint	string	validate	validate (passed validation), train (as trained), base (the weights the experiment started from), or a checkpoint directory
model.name	string | null		Instead of an experiment: a Hugging Face model ID or a local directory
model.revision	string | null		40-character commit of name; required unless it is a directory
model.architecture	string | null		With name: the model catalog entry (e.g. bi_encoder) for recipes that load the model themselves
data	EvaluationData	required
data.sources	list[Keyed]	required	Rows in the format of the experiment's recipe
data.sources.name	string	required

evaluation_embedding

Key	Type	Default	Description
tasks	list[string]	required	MTEB task names, e.g. STSBenchmark, NFCorpus, Banking77Classification
max_length	integer | null		Tokens per text; null: the experiment's training.max_length
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
model	SourceModel	required
model.experiment	string | null		Experiment directory (output/-) with its config.yaml
model.checkpoint	string	validate	validate (passed validation), train (as trained), base (the weights the experiment started from), or a checkpoint directory
model.name	string | null		Instead of an experiment: a Hugging Face model ID or a local directory
model.revision	string | null		40-character commit of name; required unless it is a directory
model.architecture	string | null		With name: the model catalog entry (e.g. bi_encoder) for recipes that load the model themselves

evaluation_wakeword

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
evaluation	EvaluationSettings
evaluation.batch_size	integer	8
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
model	WakewordModel	required
model.experiment	string | null		A wakeword experiment directory (output/-)
model.checkpoint	string	validate	validate or train (a stage of the experiment), or a bundle directory: wakeword.json beside the detector and the two feature models
data	WakewordEvaluationData	required
data.positive	list[Keyed]	required	Recordings of the phrase (audio, optional text), ideally by voices the detector never trained on
data.positive.name	string	required
data.negative	list[Keyed]		Short recordings that must not wake it (near misses, commands); the report counts them per text
data.negative.name	string	required
data.ambient	list[string]		Directories of WAVs that never say the phrase (speech, TV, music), hours of them: false accepts per hour
data.noise	list[string]		Directories of noise WAVs mixed under the recordings in the noisy conditions; empty: Gaussian noise
conditions	list[number | null]		Speech-to-noise ratios in dB every recording is tested at; null: clean
threshold	number	0.5	The operating point the headline numbers are read at
refractory	number	1.0	Seconds after a firing in which another one is not counted
targets	list[number]		False accepts per hour at which the miss rate is read off the threshold sweep

llm_decision_cispo

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.head_lr	number	0.0001	Learning rate of the head, which starts untrained
method.calibration	number	0.1	Share of training records kept for fitting the temperature
method.max_think	integer	256	Reasoning tokens per question before the model must decide
method.eval_think	boolean	True
method.group_size	integer	4	Reasoning chains sampled per question
method.temperature	number	1.0
method.rollout_weight	number	0.5	Weight of the head loss after reasoning
method.anchor_weight	number	0.5	Weight of the head loss without reasoning
method.length_hinge	integer | null		Chains longer than this many tokens lose reward
method.length_penalty_cap	number	0.1
method.iterations	integer	1	Optimizer steps per sampled group of chains
method.clip	number	0.2	The importance weight of a reused chain token is capped at 1 + clip

llm_decision_sft

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.head_lr	number	0.0001	Learning rate of the head, which starts untrained
method.calibration	number	0.1	Share of training records kept for fitting the temperature
method.max_think	integer	256	Reasoning tokens per question before the model must decide
method.eval_think	boolean	False	Also measure accuracy when the model reasons first
method.think_fraction	number	0.0	Share of questions given a reasoning chain

llm_dpo

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.beta	number	0.1	How strongly the policy may move away from the base model

llm_grpo

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.group_size	integer	4	Completions sampled per prompt
method.max_new_tokens	integer	32
method.temperature	number	1.0
method.reward	Keyed
method.reward.name	string	required
method.iterations	integer	1	Optimizer steps per sampled group
method.clip	number	0.2	The policy ratio is held within 1 +- clip when a group is reused
method.beta	number	0.0	Weight of the KL penalty towards the base model; needs training.adapter

llm_instruction

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config
method.system	string | null		System prompt put in front of every instruction

llm_sft

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	TuningValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.batch_size	integer	8
model	LLMBackboneConfig	required
model.architecture	string	required	Key of the family's model catalog
model.name	string	required	Hugging Face model ID, or a local directory
model.revision	string | null		40-character commit; required unless name is a directory
model.init	string | null		Start from this checkpoint directory of an earlier experiment
model.template	object (extra keys allowed)		Extra chat-template arguments, e.g. enable_thinking
model.head	string | null		Decision head on top of the backbone; decision recipes only
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	TrainingConfig
training.epochs	integer	1
training.lr	number	1e-05
training.weight_decay	number	0.0
training.batch_size	integer	8
training.accumulation	integer	1	Batches per optimizer step
training.warmup_ratio	number	0.05
training.max_grad_norm	number	1.0
training.max_length	integer	1024	Tokens per sequence
training.overflow	error | skip	error	Rows over max_length: stop before training, or leave them out
training.max_steps	integer | null		Pilot: stop after this many optimizer steps
training.resume_every	integer | null	200	Steps between resume snapshots; null disables them
training.eval_every	integer | null		Steps between evaluations on the held-out rows while training; null: only validate at the end
training.eval_rows	integer | null		How many held-out rows the in-training evaluation uses; null: all of them
training.adapter	AdapterConfig | null		null trains every weight; set it to train LoRA
training.adapter.name	lora	lora
training.adapter.r	integer	16	Rank of the update matrices
training.adapter.alpha	integer	32	Update scale; the effective factor is alpha / r
training.adapter.dropout	number	0.0
training.adapter.targets	list[string] | null		Layer names to adapt; default: the architecture's
method	Config

tts_piper

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	PiperValidationConfig
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.asr	AsrConfig
validation.asr.name	string	openai/whisper-base	Whisper model on the Hub
validation.asr.revision	string	e37978b90ca9030d5170a5c07aadb050351a65bb	40-character commit
validation.asr.language	string | null		Language code (en, ko, ...); null lets Whisper detect it
validation.rows	integer | null	50	Held-out rows to synthesize and transcribe; null: all of them
validation.phrases	dict[str, string]		Fixed replies the agent plays (name -> text), e.g. web_search: Searching Reference.; the best-heard take is kept as phrases/.wav and must be understood
validation.phrase_takes	integer	3	Takes of each phrase; Piper's unseeded noise makes each one different, and Whisper picks the best
model	VoiceModelConfig	required
model.quality	medium | high	medium	Piper's model size; init must have the same
model.espeak_voice	string	required	espeak-ng voice that turns text into phonemes, e.g. en-us, en-gb-x-rp
model.sample_rate	integer	22050
model.init	HubFile	required	Piper checkpoint (.ckpt) whose weights training starts from
model.init.name	string	required	Hub repository ID, or a local file or directory
model.init.revision	string | null		40-character commit; with file, for a Hub repository
model.init.file	string | null		Path of the file inside the Hub repository
data	DataConfig	required
data.sources	list[Keyed]	required
data.sources.name	string	required
data.validation	number	0.1	Share of rows held out for the validate stage
training	PiperTrainingConfig
training.epochs	integer	100	Epochs over the training rows from the starting weights
training.batch_size	integer	16
training.args	dict[str, string | integer | number | boolean]		More piper.train fit options, e.g. model.learning_rate: 1.0e-4

wakeword

Key	Type	Default	Description
recipe	string	required	Which pipeline builds this model; see llmclick recipes
name	string	required
seed	integer	20260920
output_dir	string	./output
device	cuda | mps | cpu | null
stages	list[string] | null		Stages to run, in the recipe's order; default all
tracker	TrackerConfig
tracker.enabled	boolean	False
tracker.project	string	llmclick
tracker.entity	string | null
tracker.tags	list[string]
validation	WakewordValidation
validation.min	dict[str, number]		metric -> lowest accepted value
validation.max	dict[str, number]		metric -> highest accepted value
validation.threshold	number	0.5	Score at or above which the detector fires
validation.snr	number | null	20.0	Speech-to-noise ratio of the test streams, dB; null: clean
phrase	string	required	What wakes the agent, e.g. 'Jarvis, wake up'; names the model file
data	WakewordData	required
data.positive	list[Keyed]	required	Row sources of recordings of the phrase (audio, optional text)
data.positive.name	string	required
data.negative	list[Keyed]	required	Row sources of anything else: near misses, commands, speech
data.negative.name	string	required
data.background	list[string]		Directories of noise WAVs mixed under every window; empty: Gaussian noise
data.validation	number	0.2	Share of each class held out for the streaming test
augment	AugmentConfig
augment.rounds	integer	4	Windows cut from every training recording, each with its own noise
augment.snr	list[any]	[5.0, 25.0]	Range of speech-to-noise ratio, dB
augment.gain	list[any]	[-12.0, 0.0]	Range of gain on the whole window, dB
augment.tail	number	0.3	Most seconds of noise after a recording's end in its window
augment.followed	integer	2	Negative windows per phrase recording where other speech follows the phrase, so the detector fires as the phrase ends and not again
augment.noise	integer	200	Windows of noise alone, as more negatives
training	WakewordTraining
training.steps	integer	3000
training.batch_size	integer	256	Half positive windows, half negative
training.learning_rate	number	0.001
training.layer_dim	integer	32	Width of openWakeWord's DNN head
training.blocks	integer	1	Hidden blocks after the first layer
training.negative_weight	number	1.0	Loss weight of a negative window against a positive one; above 1 trades unused recall for fewer false accepts
training.schedule	constant | cosine	constant	Learning rate over the steps; cosine decays it to zero, so the detector does not depend on where training stops