Configuration reference

Every run is described by DistillationConfig (config.py). The web app builds it from the form; the CLI from a YAML/JSON file plus options. Invalid values are reported with the field name.

This page is checked against the model by tests/test_docs.py.

Field	Type	Default	Allowed	Meaning
teacher_model	str	required		Model name as the endpoint knows it; comma-separate several for an ensemble.
judge_model	str \| None	None		Grades pairs in Balanced and Research mode. Default: the first teacher.
judge_threshold	int	4	≥ 1 ≤ 5	Minimum judge score (1-5) that faithfulness, helpfulness and correctness must all reach.
embedding_model	str \| None	None		Embedding model on the same endpoint; enables paraphrase removal.
semantic_dedup_threshold	float	0.92	≥ 0.5 ≤ 0.999	Cosine similarity at which a pair counts as a paraphrase of an accepted one.
base_url	str \| None	None		OpenAI-compatible endpoint, e.g. http://localhost:8000/v1. Unset: OpenAI.
dataset_size	int	500	≥ 10 ≤ 50000	Target number of pairs.
quality_mode	QualityMode	'balanced'		fast (filters), balanced (+ judge) or research (+ evolved questions).
output_format	OutputFormat	'alpaca'		alpaca, sharegpt, chatml or openai.
train_model	bool	False		Train a LoRA adapter after export (needs the train extra).
base_model	str	'Qwen/Qwen3-4B-Instruct-2507'		Hugging Face model id to fine-tune.
publish_dataset	bool	False		Upload the dataset to the Hugging Face Hub with a dataset card.
hf_repo	str \| None	None		Dataset repo, user/name.
hf_private	bool	True		Create repos as private.
dataset_license	one of ['other', 'unknown', 'cc-by-4.0', 'cc-by-sa-4.0', 'cc-by-nc-4.0', 'cc0-1.0', 'odc-by', 'mit', 'apache-2.0']	'other'		License written on the dataset and model cards.
publish_adapter	bool	False		Also upload the LoRA adapter as a model repo (needs train_model).
hf_model_repo	str \| None	None		Adapter repo. Default: <hf_repo>-lora.
temperature	float	0.7	≥ 0.0 ≤ 2.0	Sampling temperature for questions and answers (the judge always uses 0).
max_new_tokens	int	2048	≥ 128 ≤ 32768	Maximum tokens per answer.
concurrency	int	8	≥ 1 ≤ 64	Parallel requests per model client. Use 1-2 for slow CPU servers.
request_timeout	int	120	≥ 10 ≤ 1800	Seconds before a request is retried.
seed	int \| None	None	≥ 0 ≤ 2147483647	Sampling seed sent with every request. Unset: a random seed is chosen and recorded. A server that rejects the field gets requests without it (logged).
reasoning_effort	none, minimal, low, medium, high or None	None		Thinking effort for models that think. none turns thinking off where the server allows it (Gemini 2.5 Flash), so it does not use up the answer's token budget. Unset: the server's default.
lora_rank	int	16	≥ 4 ≤ 256	LoRA rank r (alpha is set equal).
api_key	str \| None	None		Key for the endpoint. Never written to manifests or logs. CLI: environment only.
hf_token	str \| None	None		Hugging Face write token. Same rules as api_key.
use_semantic_chunking	bool	False		Paragraph and sentence-aware chunks instead of fixed windows.
enable_dedup	bool	True		Remove near-duplicates (and paraphrases when an embedding model is set).
sanitize_dataset	bool	False		Redact PII, strip HTML and apply quality gates before export.
pii_url_policy	one of ['domain', 'redact', 'keep']	'domain'		domain keeps scheme and host, redact removes links, keep keeps them. Credentials and secret query values are always removed.
pii_presidio	bool	False		Also detect names and ID numbers with Presidio (pii extra).
decontaminate	list[str]	[]		Benchmarks to remove overlap with: gsm8k, mmlu, arc, truthfulqa, humaneval.

Rules across fields

• The OpenAI API (no base_url) needs an api_key; local servers do not.
• Publishing needs hf_repo and hf_token.
• publish_adapter needs train_model, a repo (hf_model_repo or hf_repo) and a token.

Environment

See the operator table in the README: OPENAI_BASE_URL, OPENAI_API_KEY, HF_TOKEN, BRAINBREW_ALLOW_CUSTOM_ENDPOINTS, BRAINBREW_DEFAULT_MODEL, BRAINBREW_MAX_JOBS, BRAINBREW_RUNS_DIR, BRAINBREW_LOG_FORMAT, BRAINBREW_LOG_LEVEL, BRAINBREW_REQUIRE_LOGIN.