`interlens.templating`¶

interlens.templating ¶

Per-row templating for data-driven rollouts.

A Conversation run over a benchmark shares ONE recipe but needs a DIFFERENT scenario per dataset row. Any templated string field on the conversation (shared_context, shared_system_prompt) or its participants (system_prompt, private context) may be a template value that is resolved against each row at rollout expansion time (once, deterministically — so job i always maps to row i and resume stays correct).

A template value is any of

a plain str — used as-is (no substitution);
dataset_field("question") — pulls row["question"];
a callable(row) -> str — arbitrary per-row computation;
a tuple/list of the above — resolved part-by-part and concatenated (the <3.14 way to interleave literal text with fields: ("Solve:\n\n", dataset_field("question")));
a PEP 750 template string (t"...{dataset_field('question')}...") on Python 3.14+ — interpolated DatasetFields resolve from the row, everything else formats normally.

resolve(value, row) performs the substitution; has_fields(value) reports whether a value depends on the row at all (so a conversation with no data-dependent fields can be run directly, and one WITH them refuses to run un-expanded — see Conversation).

DatasetField `dataclass` ¶

DatasetField(name: str)

A placeholder for row[name] in a templated field, created by :func:dataset_field. Frozen + tiny so it serializes and pickles trivially across a spawn boundary.

dataset_field ¶

dataset_field(name: str) -> DatasetField

Reference a dataset column in a templated field: shared_context=("Solve:\n\n", dataset_field("question")) resolves row["question"] for each row at rollout expansion. See the module docstring for all template forms.

Source code in src/interlens/templating.py

def dataset_field(name: str) -> DatasetField:
	"""Reference a dataset column in a templated field: ``shared_context=("Solve:\\n\\n", dataset_field("question"))``
	resolves ``row["question"]`` for each row at rollout expansion. See the module docstring for all template forms."""
	return DatasetField(name)

has_fields ¶

has_fields(value) -> bool

Whether value depends on the dataset row (contains a DatasetField, a callable, or a t-string). A conversation whose framing has no fields can be run directly; one that has them must be expanded over data.

Source code in src/interlens/templating.py

def has_fields(value) -> bool:
	"""Whether ``value`` depends on the dataset row (contains a ``DatasetField``, a callable, or a t-string). A
	conversation whose framing has no fields can be run directly; one that has them must be expanded over data."""
	if isinstance(value, DatasetField) or callable(value):
		return True
	if isinstance(value, (tuple, list)):
		return any(has_fields(v) for v in value)
	if _Template is not None and isinstance(value, _Template):
		return True
	return False

resolve ¶

resolve(value, row: dict)

Resolve a template value against one dataset row (a mapping) to a concrete value.

Non-templated values (plain str/None/other) pass through unchanged; DatasetField → row[name]; a callable → value(row); a tuple/list → each part resolved and string-joined; a t-string → interpolated with DatasetField interpolations pulled from the row.

Source code in src/interlens/templating.py

def resolve(value, row: dict):
	"""Resolve a template ``value`` against one dataset ``row`` (a mapping) to a concrete value.

	Non-templated values (plain ``str``/``None``/other) pass through unchanged; ``DatasetField`` → ``row[name]``;
	a callable → ``value(row)``; a tuple/list → each part resolved and string-joined; a t-string → interpolated
	with ``DatasetField`` interpolations pulled from the row."""
	if isinstance(value, DatasetField):
		return row[value.name]
	if _Template is not None and isinstance(value, _Template):
		return _resolve_template_string(value, row)
	if isinstance(value, (tuple, list)):
		return "".join(str(resolve(v, row)) for v in value)
	if callable(value):
		return value(row)
	return value

interlens.templating¶