bake.bake¶

bake.bake(
    what,
    data=None,
    *,
    method='auto',
    student=None,
    teacher=None,
    labels='auto',
    where='auto',
    test=None,
    metric=None,
    report=True,
    compare_teacher=False,
    wait=True,
    plan_only=False,
    path=None,
    run_folder=None,
    fixed=None,
    derived=None,
    layout=None,
    reasoning=False,
    tags=None,
    weights=None,
    functions=None,
    name=None,
    validation=None,
    holdout=None,
    num_threads=16,
    local_files_only=False,
    log=True,
    seed=0,
    **training,
)

Train weights that answer an AI function (or several); returns the baked model.

what and data:

• fn, rows: one function; rows are dicts (or any table dpyr.read takes) with the inputs under the parameters' names and, when known, the answers under the outputs' names (result for the return value).
• {fn: rows, fn2: rows2}: one student for several functions, each called through its own layout.
• program, rows: a @module; it runs on each row (with teacher answering every AI call inside it), and every call becomes an example of its function (functions= keeps some).
• examples: made by functai.bake.examples (or a file of them).

The main choices (all decided from the data when left out; plan_only=True prints the plan and spends nothing):

• method: "head", "sft", or "auto" (a head when every output is finite and nothing asks for a generative student).
• student: a Hugging Face model id or folder.
• teacher: answers the rows without answers: a model name, an AI function, or {fn: teacher}; default each function's own model. labels="teacher" asks it for every row; "data" uses only rows with answers.
• where: "here", "tinker", "prime", "export", a list in order of preference, or a Trainer; "auto": here when this machine can train it, else the cheapest service set up (configure(bake_where=...) sets a preference once).
• fixed={"input": value}: an input with one value in every row, left out of the student's prompt; a call with another value is refused. derived={"input": "other input"}: an input decided by another, left out too.
• test: rows to judge on (else a share of the rows with answers is set aside); metric: how to score them (anything evaluate takes; an AI judge for open text); report=False skips judging.
• wait=False: return the Run at once (a folder and a process that outlive this one); running the same bake again resumes it.
• training settings: lora (True/False), lora_rank, lr, epochs, batch (examples per step), quantize="4bit", packing, devices, liger, max_new_tokens, merge (merge the adapter into the weights; default True), report_to (["wandb"], ...).
• for a head: epochs, lr, batch_size, max_length, device, prices, holdout, validation as before.