bake.Baked¶

bake.Baked(path, *, device=None, check=True)

A model trained to answer one or more AI functions: what bake returns, and what functai.bake.load(folder) reads back.

Use it as a model: fast = summarize.using(lm=baked) is the same function, answered by the weights (fast("..."), fast.map(rows), functai.evaluate(fast, rows)). Calls send exactly the tokens the student was trained on.

Two kinds (kind): a head ("head") answers functions whose every output has a fixed set of answers, with a probability for each (probabilities, predict for many rows at once); a generative student ("generative", trained with method="sft") writes its answer like any chat model, and runs in this process, on a vLLM server, on Tinker, or at any OpenAI-compatible address (on).

On disk it is a folder: baked.json (what it answers: per function, its signature, its layout and the inputs left out; the weights' form and base model; the chat template; the run that made it; the report; file hashes, checked when loading), model/ (Hugging Face weights, merged when trained with LoRA: a standard folder for vLLM, TGI or transformers), adapter/ (the LoRA adapter alone), tokenizer/, and heads.safetensors for a head of several outputs. Weights trained on a service and not brought here yet are a tinker:// address (download() brings them here).

Attributes¶

Name	Description
`capabilities`	What its calls can do (tools, streaming...), as lmcc lays out requests for it.
`device`	Where it runs in this process (`cuda`, `mps` or `cpu`): the `device` given to
`endpoint`	The address calls are sent to when it runs on a server (`on("vllm")`, a URL); else None.
`fingerprint`	The fingerprint of the signature it was trained to read (one function).
`functions`	The names of the AI functions it answers.
`layout`	The lmcc adapter artifact its function's calls are written in (one function).
`model`	The model name its calls are logged under: `baked:<name>`.
`name`	The model's name (by default the function's, or the functions' joined by `+`).
`provider`	The provider name its calls are logged under.
`report`	How it did on its test rows when it was baked (or after `judge(..., save=True)`); None
`runner`	What answers its calls now (in this process by default; see `on`).
`signature`	What the student reads (one function).
`student`	The base model it was trained from (a Hugging Face id).
`template`	The chat template it was trained with (its text and hash): calls must write the same.

Methods¶

Name	Description
call_signature	The signature `fn`'s calls bind on this model: what the student reads.
complete	Answer one lm15 request (what FunctAI calls; to call the function on it, use
download	Bring weights trained on a service here (merged into a standard
entry_for	The entry for `fn` (refused when the model was not trained for it,
on	Run on `where`: `"transformers"` (in this process), `"vllm"` (a
predict	A head model's answers for many rows at once (the fast path for big
probabilities	Per text, the probability of every answer, per field (texts as the layout writes them).
reduce	(spec, inputs) as the student reads them (fixed and derived inputs left
requirements	The packages running it needs here.
resolve	Where a call named `model` goes (used by FunctAI when it routes a call).
save	Copy the model to `path`; returns it loaded from there.
serve	Serve with vLLM and send calls there; returns the endpoint.
size	Its folder's size on disk, in bytes.
stop	Stop a server this model started (`serve()`/`on("vllm")`); calls run in-process again.
texts	The input text the model reads for each row of inputs (written by its layout).
tokenizer	Its tokenizer, checked to carry the chat template it was trained with.

call_signature¶

bake.Baked.call_signature(fn, spec)

The signature fn's calls bind on this model: what the student reads.

complete¶

bake.Baked.complete(request)

Answer one lm15 request (what FunctAI calls; to call the function on it, use fn.using(lm=baked)).

download¶

bake.Baked.download(path=None)

Bring weights trained on a service here (merged into a standard folder). Returns the model, loaded from its folder.

entry_for¶

bake.Baked.entry_for(fn, spec)

The entry for fn (refused when the model was not trained for it, or when it changed since).

on¶

bake.Baked.on(where=None, **options)

Run on where: "transformers" (in this process), "vllm" (a server started here), "tinker", or an OpenAI-compatible URL. Returns the model.

predict¶

bake.Baked.predict(rows)

A head model's answers for many rows at once (the fast path for big tables): per row, each field's answer, its probability, and the full distribution.

probabilities¶

bake.Baked.probabilities(texts)

Per text, the probability of every answer, per field (texts as the layout writes them).

reduce¶

bake.Baked.reduce(fn, spec, inputs, *, check=True)

(spec, inputs) as the student reads them (fixed and derived inputs left out, after checking their values).

requirements¶

bake.Baked.requirements()

The packages running it needs here.

resolve¶

bake.Baked.resolve(model)

Where a call named model goes (used by FunctAI when it routes a call).

save¶

bake.Baked.save(path, *, overwrite=False)

Copy the model to path; returns it loaded from there.

serve¶

bake.Baked.serve(**options)

Serve with vLLM and send calls there; returns the endpoint.

size¶

bake.Baked.size()

Its folder's size on disk, in bytes.

stop¶

bake.Baked.stop()

Stop a server this model started (serve()/on("vllm")); calls run in-process again.

texts¶

bake.Baked.texts(rows)

The input text the model reads for each row of inputs (written by its layout).

tokenizer¶

bake.Baked.tokenizer()

Its tokenizer, checked to carry the chat template it was trained with.