bake.Examples¶

bake.Examples(rows, *, student=None, template=None, entries=None, info=None)

The training conversations for a student, made by functai.bake.examples: what FunctAI will send the student, in a form every trainer reads.

One row per example: function (the AI function it trains); messages (the chat as the function's layout writes it, the reply last: TRL, Axolotl, Unsloth and prime-rl read it as is); prompt and completion (the same chat split at the reply, written on save); with student=, input_ids (the exact tokens under the student's chat template: what training sees and what a call sends), answer_start (where the reply starts: the loss is on the tokens from there on), prompt_tokens and answer_tokens; tag and weight (from your columns); row_id and split (train or validation); source (data, teacher or program).

ex = functai.bake.examples(summarize, rows, student="Qwen/Qwen3.5-4B")
ex.stats()                                   # counts and token lengths
ex = ex.filter(lambda r: r["prompt_tokens"] < 16_000)
ex.save("train.parquet")                     # or .jsonl / .jsonl.gz; ex.to_hf() for datasets
baked = functai.bake.bake(ex)                # or train elsewhere, then functai.bake.adopt(...)

Attributes¶

Name	Description
`functions`	The names of the AI functions the examples train.
`table`	The examples as a dpyr table (`pip install "functai[data]"`).
`tokenized`	Whether every example carries its tokens (`input_ids`: made with `student=`).

Methods¶

Name	Description
filter	The examples `keep(row)` is true for: `ex.filter(lambda r: r["prompt_tokens"] < 16_000)`.
from_records	Examples from plain rows (each needs `function` and `messages`), as `records()` gives
load	Examples written by `save` (or by a bake run).
records	Plain rows for files and tables: `prompt` and `completion` added,
save	Write the examples: `.parquet`, `.jsonl` or `.jsonl.gz`. What
split	The examples of one split: `ex.split("train")` or `ex.split("validation")`.
stats	Counts and token lengths (lengths need tokens: `student=`).
to_hf	A Hugging Face `datasets.Dataset`.

filter¶

bake.Examples.filter(keep)

The examples keep(row) is true for: ex.filter(lambda r: r["prompt_tokens"] < 16_000).

from_records¶

bake.Examples.from_records(recs, meta=None)

Examples from plain rows (each needs function and messages), as records() gives them; meta carries the student, template and layouts when known.

load¶

bake.Examples.load(path)

Examples written by save (or by a bake run).

records¶

bake.Examples.records(columns=None)

Plain rows for files and tables: prompt and completion added, token ids as lists.

save¶

bake.Examples.save(path, *, columns=None)

Write the examples: .parquet, .jsonl or .jsonl.gz. What made them (the functions' layouts, the student's template) goes in the Parquet file's metadata, or a .meta.json file beside a JSONL one.

split¶

bake.Examples.split(name)

The examples of one split: ex.split("train") or ex.split("validation").

stats¶

bake.Examples.stats()

Counts and token lengths (lengths need tokens: student=).

to_hf¶

bake.Examples.to_hf(columns=None)

A Hugging Face datasets.Dataset.