Config file

config.yaml controls the pipeline through five sections: base, tuning, evaluation, synthgen and trace_processing. Only base.task is required, and every other field has a default. Always set base.task, base.student_model_name and base.teacher_model_name.

File format

The config is YAML, in a file named config.yaml (config.yml is accepted too). The CLI reads it out of your data directory, and the REST API stages it as config_yaml.

base:
  task: question-answering
  student_model_name: Qwen3.5-4B
  teacher_model_name: zai.glm-5.3-low-thinking
evaluation:
  llm_as_a_judge_model_name: zai.glm-5.3-low-thinking

Overrides sent inline over the REST API carry this same structure in the request body. See Using the REST API.

base

Parameter	Default	Notes
task	required	The task the student solves: question-answering, classification, chat-completion or chat-completion-agentic. See Task selection
visual_task	false	The task inputs carry images. Question answering only; every model must be vision-capable. Image traces use the openai_messages_with_images observation format
enable_thinking	false	The student is trained, evaluated and served with its chat template’s thinking mode on. Synthetic data generation then writes the reasoning: the teacher produces it with each assistant message it generates, and a final backfill pass fills in what is still missing. Reasoning students only. See Reasoning models
variable_system_prompt	false	false: every example is trained, evaluated and served with the task_description as its system message, which replaces a system message found in a row or trace. true: every row and relabeled trace keeps its own system message (a row without one gets the task_description), and the served model reads the system message each request sends. See System prompts
student_model_name	Llama-3.2-1B-Instruct	The base model that is fine-tuned into your student. See Supported models
teacher_model_name	openai.gpt-oss-120b	The model that generates the synthetic data and that the student learns from. See Supported models
random_seed	123	Seed for random sampling across the pipeline, mutators included
llm_num_parallel_requests	4	Maximum number of LLM requests sent in parallel, across the teacher, synthetic generation and the judge. 1 turns parallelism off
should_expand_dataset	"auto"	Whether a conversation becomes one example per assistant turn: true, false or auto. See Conversation expansion

tuning

Parameter	Default	Notes
learning_rate	5e-5	Initial learning rate of the AdamW optimizer
learning_rate_scheduler	linear	cosine, linear or constant. See the Hugging Face SchedulerType for all values
weight_decay	0.0	AdamW weight decay, applied to all layers except the bias and LayerNorm weights
warmup_ratio	0.05	Fraction of the training steps used for a linear warmup from 0 to learning_rate
bf16	true	bf16 mixed-precision training instead of 32-bit training
lora_r	64	Rank of the LoRA adapter; students are always trained with LoRA. One of the ranks vLLM serves LoRA adapters at: 8, 16, 32, 64, 128, 256, 320 or 512
lora_alpha_multiplier	1	LoRA alpha = lora_r * lora_alpha_multiplier
per_device_train_batch_size	1	Batch size per device for fine-tuning
per_device_eval_batch_size	1	Batch size per device of the trainer’s evaluation on the validation set. Used only with enable_trainer_internal_eval: true, never by the post-training evaluation
num_train_epochs	4	Number of fine-tuning epochs
gradient_accumulation_steps	1	Steps whose gradients accumulate before one optimizer update. Multiplies the effective batch size without using more memory
enable_trainer_internal_eval	false	Fine-tuning and RLVR evaluate on the validation set, which is the test split, before training and at the end of every epoch. Final metrics come from the post-training evaluation either way; false trains faster and with less memory
logging_steps	0.1	How often fine-tuning and RLVR log their metrics, as a fraction of all optimizer steps (0.01 logs every 1%). Above 0, at most 1
save_steps	0.25	How often fine-tuning and RLVR save a checkpoint, as a fraction of all optimizer steps (0.1 saves every 10%). Above 0, at most 1
memory_optimized_training	0	0, 1 or 2; raise it only when training runs out of GPU memory. 1: gradient checkpointing, which recomputes activations in the backward pass instead of storing them, and is slower (RLVR also puts vLLM to sleep between generations). 2: 1 plus activation offloading to CPU RAM, much slower. true reads as 2 and false as 0. See Out-of-memory failures
use_qlora	false	Loads the base model in 4-bit NF4 (QLoRA) for fine-tuning and attaches the LoRA adapter in higher precision. The base model uses about a third of the GPU memory, and training is slightly slower
max_completion_length	2048	Tokens the student may generate: caps generation in student evaluation and the rollouts of RLVR. Teachers use synthgen.teacher_max_tokens instead. Raise it when training warns it is below the longest training completion
rlvr_dataset_size	0.0	Fraction of the dataset in the RLVR split, from 0 to 1. RLVR, an optional reinforcement learning stage after fine-tuning, runs when this is above 0
rlvr_llm_as_a_judge_model_name	inherits base.teacher_model_name	The judge that scores RLVR rollouts
rlvr_reward_metric	llm-judge-reference-free	How the judge scores a rollout. llm-judge-reference-free: against the problem alone. llm-judge-reference: the judge also sees the reference answer. llm-judge-reference-without-context: the judge sees the reference but not the problem, which is much cheaper on long problems and enough when the reference fixes the answer exactly
rlvr_per_device_batch_size	6	Completions per forward and backward pass in RLVR training and evaluation. rlvr_per_device_batch_size * rlvr_gradient_accumulation_steps must be a multiple of rlvr_num_generations, so the gradients from one prompt stay in one optimizer step
rlvr_num_generations	6	Completions sampled per prompt in RLVR training
rlvr_gradient_accumulation_steps	1	Micro-batches per RLVR optimizer step, separate from gradient_accumulation_steps
rlvr_num_train_epochs	1	Number of RLVR epochs
rlvr_learning_rate	5e-5	Initial AdamW learning rate of RLVR
rlvr_weight_decay	0.0	AdamW weight decay of RLVR, applied to all layers except the bias and LayerNorm weights
rlvr_learning_rate_scheduler	constant	Learning rate scheduler of RLVR; learning_rate_scheduler does not apply to it
rlvr_scale_rewards	batch	How GRPO turns the centred rewards into advantages. group divides by the standard deviation of each prompt’s group, which turns judge noise in an otherwise uniform group into a full-size advantage. batch divides by the standard deviation of the whole batch, so the advantage stays proportional to the reward gap. none leaves the centred rewards unscaled
rlvr_vllm_importance_sampling_mode	token_truncate	How GRPO corrects for vLLM and the trainer giving slightly different log-probabilities to the same rollout. token_* uses one ratio per token, sequence_* one ratio per completion (the product over its tokens, which drifts far from 1 on long completions). *_truncate clips the ratio at rlvr_vllm_importance_sampling_clip_max, *_mask zeroes the loss above it
rlvr_vllm_importance_sampling_clip_max	2.0	Upper bound of the vLLM importance-sampling ratio; null leaves it unbounded
rlvr_steps_per_generation	null	Micro-batches of rlvr_per_device_batch_size rollouts generated in one go. null means rlvr_gradient_accumulation_steps: each generation feeds exactly one optimizer step, so training is on-policy and rlvr_epsilon and rlvr_epsilon_high never bind. A multiple k of rlvr_gradient_accumulation_steps spreads one generation over k optimizer steps (GSPO): the steps after the first train on rollouts from a slightly older policy, and the clip bounds how far each of them moves it
rlvr_importance_sampling_level	sequence	Level of the ratio between the current policy and the policy that generated the rollouts, which rlvr_epsilon and rlvr_epsilon_high clip. token: one ratio per token (GRPO). sequence: one length-normalised ratio per completion (GSPO). Matters only when rlvr_steps_per_generation makes training off-policy
rlvr_epsilon	3e-4	Lower clip of that ratio: an update may lower a rollout’s ratio to at most 1 - rlvr_epsilon. The default suits sequence-level ratios (GSPO); token-level GRPO usually uses 0.2
rlvr_epsilon_high	4e-4	Upper clip of that ratio: an update may raise a rollout’s ratio to at most 1 + rlvr_epsilon_high. The default suits sequence-level ratios (GSPO); token-level GRPO usually uses 0.28
rlvr_overlong_punishment_cache	null	Length penalty for RLVR rollouts (DAPO soft overlong punishment): the number of tokens before max_completion_length over which the penalty ramps up. A rollout of up to max_completion_length - rlvr_overlong_punishment_cache tokens has no penalty; above that the penalty falls linearly from 0 to -1 at max_completion_length. For example, max_completion_length 6144 and rlvr_overlong_punishment_cache 2048 give no penalty up to 4096 tokens, -0.5 at 5120 and -1 at 6144. The penalty is added to the judge score and also applies when the judge could not score the rollout. Must be below max_completion_length; null turns it off
rlvr_vllm_gpu_memory_utilization	0.3	Fraction of the GPU the vLLM engine that generates the rollouts may use for its copy of the student and its KV cache. The training copy of the student uses the rest, so a student whose weights alone exceed this fraction cannot start its vLLM engine

evaluation

Parameter	Default	Notes
llm_as_a_judge_model_name	inherits base.teacher_model_name	The model behind the LLM-as-a-judge. Scores from different judges can’t be compared. See Supported models

synthgen

Parameter	Default	Notes
train_generation_target	10000	Synthetic rows one generate-synthetic-data-train run adds to the train split; 0 generates nothing. A target, not an exact count: generation runs in generation_iteration_size batches until the target is met, so the result can exceed it by up to one batch
test_generation_target	0	Synthetic rows one generate-synthetic-data-test run adds to the test split; 0 generates nothing. Same rounding
use_traces_as_context	true	The traces in the dataset serve as context for generation and the used ones leave the dataset: T + min(T, 1000) for a target of T, or all that are left. Below min(T / 4, 10) traces the job logs a warning and generates without context; with false it generates without context. Either way it leaves the traces alone. See The expand operations
generation_in_single_call	4	Examples generated per teacher call
generation_iteration_size	128	Examples per generation batch: generation and validation run in batches of this size
num_positive_exemplars_per_generation	1	In-context examples per generation call of the class or task being generated
num_negative_exemplars_per_generation	1	In-context examples of the classes not being generated. Classification only
num_unlabelled_exemplars_per_generation	1	Unlabeled examples (context traces) shown per generation call, when the run has context
validation_max_total_length	30000	Maximum total length of an example, input plus output, in characters. Applies to uploaded data too
validation_similarity_threshold	0.95	Generated rows whose similarity to the rows already in the dataset is above this threshold are removed. Lower it if generation produces near-duplicates
validation_instructions	null	Instructions for an LLM validator of synthetic examples. The evaluation judge (by default the teacher) checks each complete example, system prompt included, against these instructions and the job description, after the structural and diversity checks. Rejected examples and invalid judge replies are dropped. null turns it off and makes no validation LLM calls. See LLM validation
teacher_temperature	0.7	Sampling temperature of the teacher, from 0 to 1: higher gives more varied output. See Cross-field validation for reasoning teachers
teacher_max_tokens	32000	Maximum tokens in one teacher response. Kept well below typical context limits, so the reserved output does not crowd out large (for example multi-image) prompts
match_generated_distribution_to_seed	false	Generated data keeps the class distribution of the seed rows. Classification only
output_is_json	false	Question answering only: only synthetic rows whose output is valid JSON are generated. Answers in uploaded data must also be valid JSON
mutators	[]	Mutators applied during synthetic generation, one per dimension to vary. See Relabeling and synthetic data
mutator_update_frequency	5	Generation batches between two runs of the topic classifier, for every adaptive mutator. The batches in between are not classified; their counts are estimated from the classified ones. Classifying costs teacher calls, so a higher number is cheaper and coarser
max_tool_calls_per_turn	null (1 with tools declared, 0 without)	Chat completion only: the most tool calls one assistant message may contain. unlimited allows any number of parallel calls; 0 disallows tool calls. A job description without tools needs 0 or null. See Chat completion
clean_training_targets	false	A final teacher pass that minimally repairs corrupted or truncated final assistant messages (the training targets). Expansion before it follows base.should_expand_dataset; with false only the supplied final-turn targets are cleaned

trace_processing

Read by relabel-traces-train and relabel-traces-test (Relabeling and synthetic data); observation_format is also read when a dataset is created from traces.

Parameter	Default	Notes
relabel	true	The teacher (or committee) relabels the traces. false keeps the original trace labels
relevance_filtering	false	An LLM scores each trace and drops those below min_relevance_score or min_coherence_score. false skips the step and keeps every trace
relevance_filtering_batch_size	32	Traces scored per batch during relevance filtering
min_relevance_score	4	Minimum relevance score, 1-5, for a trace to pass relevance filtering
min_coherence_score	3	Minimum coherence score, 1-5, for a trace to pass. A lower value lets more corrupted traces through for committee repair
num_train_relabelled	200	Traces one relabel-traces-train run turns into train rows. The dataset must hold at least this many distinct traces or the job refuses to run; fewer rows can come out, because relabeling drops some traces. The used traces leave the dataset
num_test_relabelled	200	The same for relabel-traces-test and the test split
observation_format	openai_messages	Format of the traces in traces.jsonl. openai_messages: objects with a messages array of chat completion messages. langfuse: Langfuse observation objects (id, input, output). See Traces
remove_system_prompt_from_traces	true	Deprecated, no effect. base.variable_system_prompt decides whether traces keep their system message
compress_job_description	false	The teacher compresses the job description before relevance filtering. For a task_description so long it would overwhelm the filtering LLM
teacher_model_name	inherits base.teacher_model_name	Filters the traces and picks the best committee answer
relabelling_committee_models	[]	A non-empty list turns on committee relabeling: every model writes a candidate relabel of every trace, and teacher_model_name combines them into the final one
committee_max_input_length	250000	Characters. A trace whose combining input would exceed this skips the committee and is edited by the teacher directly. That input is the trace’s length times (1 + committee size): the original trace plus one full rewrite per member

Cross-field validation

These fail when the config is loaded.

• A reasoning teacher (every teacher except Qwen3-235B-A22B-Instruct-2507, Qwen3-480B-A35B-Coder and Qwen2.5-VL-72B-Instruct) requires synthgen.teacher_temperature between 0.5 and 0.7.
• base.enable_thinking: true requires a reasoning student.
• visual_task: true requires a vision-capable model in every teacher and judge role, and a vision-capable student (Supported models).
• Chat completion with tools declared restricts the student and teacher (Tool-calling compatibility).
• synthgen.max_tool_calls_per_turn must match whether the job description declares tools.

The two in-context exemplar counts (synthgen.num_positive_exemplars_per_generation and synthgen.num_negative_exemplars_per_generation) aren’t validated: every generation job lowers them to the number of rows the split it draws from holds, down to zero on an empty split.

Conversation expansion

base.should_expand_dataset decides whether a conversation becomes one example per assistant turn. It applies to training, to evaluation, and to the expansion before synthgen.clean_training_targets and the reasoning backfill.

Value	Effect
auto (default)	Expands multi-turn tasks, unless at least half of the examples share history prefixes with other examples (data that is already split per turn)
false	Each supplied row stays one example, with only its final assistant turn as the target; earlier assistant turns are history. Evaluation scores each test row’s final turn only. Use it for pre-split rows whose final responses were selected or rewritten
true	Expands every conversation, pre-split data included. Valid only for the chat completion tasks

Parameters that have no effect

These appear in every expanded config the platform returns and do nothing: evaluation.batch_size, synthgen.validation_max_answer_length, synthgen.parallel_llm_calls, synthgen.basic_mutators_to_use, tuning.use_lora, tuning.awq_quantize_tuned_model and tuning.train_eval_split. Leave them untouched in a config you override, and don’t add them to one you write.

evaluation.num_few_shot_examples and tuning.num_few_shot_examples_student have no effect either: teacher evaluation, training and student evaluation use each row as it is, with no examples added. To show the model examples, put them in the system prompt: task_description, or each row’s system message with base.variable_system_prompt on.

Deprecated keys

These keys are not part of the config. Each one has a replacement or no equivalent:

Deprecated key	Use instead
synthgen.generation_target	synthgen.train_generation_target
trace_processing.num_traces_as_training_base	trace_processing.num_train_relabelled
traces_to_test_set.num_traces_to_relabel	trace_processing.num_test_relabelled
traces_to_test_set.num_synthetic_examples	synthgen.test_generation_target
traces_to_test_set.evaluate_original_model	a teacher evaluation with the production model as base.teacher_model_name
traces_to_test_set.min_relabelled_examples	none
trace_processing.max_unstructured, num_traces_as_testing_base, min_generated_examples, evaluate_original_model, convert_to_single_turn	none
synthgen.generation_per_unstructured_context, num_distractor_context_blocks	none

Next

An override replaces this file whole rather than merging into it, so read the parent’s config back before you edit one field. See How the platform works.