Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Helpers for selecting the latest rows from evaluation results."""
from __future__ import annotations
from collections.abc import Iterable, Mapping, Sequence
from typing import Any
def _row_order(row: Mapping[str, Any]) -> tuple[str, int]:
"""Return the database ordering used to identify the newest row."""
created_at = row.get("created_at")
stamp = "" if created_at is None else str(created_at)
raw_id = row.get("rowid", row.get("id", -1))
try:
insertion_id = int(raw_id)
except (TypeError, ValueError):
insertion_id = -1
return stamp, insertion_id
def newest_eval_rows(
rows: Iterable[Mapping[str, Any]],
*,
key_fields: Sequence[str] = ("benchmark",),
) -> list[dict[str, Any]]:
"""Keep the newest scored row for each evaluation key.
Rows are ordered by ``created_at DESC, rowid DESC``. SQLite exposes the
``INTEGER PRIMARY KEY`` column as both ``id`` and ``rowid``; accepting both
names keeps this helper usable with database rows and small test doubles.
Rows with missing or empty key fields are ignored, matching the eval gate's
previous ``if not name`` behavior.
"""
newest: dict[tuple[Any, ...], dict[str, Any]] = {}
ordered = sorted(rows, key=_row_order, reverse=True)
for row in ordered:
if row.get("score") is None:
continue
key = tuple(row.get(field) for field in key_fields)
if any(value is None or value == "" for value in key) or key in newest:
continue
newest[key] = dict(row)
return list(newest.values())