A straightforward method for training your LLM, from downloading data to generating text.
"""
Lightweight metrics logging shared by every training stage.
Always writes one JSON object per logged step to a JSONL file under the log dir (so runs
are reproducible and plottable without any external service). Optionally mirrors to
Weights & Biases when ``use_wandb=True`` and the package is importable. Only the main
process should construct/use a logger (the trainers guard this).
"""
from __future__ import annotations
import json
import os
import time
from typing import Any
class MetricsLogger:
def __init__(
self,
stage: str,
log_dir: str,
*,
use_wandb: bool = False,
wandb_project: str = "train-llm-from-scratch-posttrain",
config: dict | None = None,
run_name: str | None = None,
) -> None:
os.makedirs(log_dir, exist_ok=True)
self.stage = stage
# No Date.now in scripts here; use time.time() (allowed) for unique filenames.
stamp = int(time.time())
self.path = os.path.join(log_dir, f"{stage}_{stamp}.jsonl")
self._fh = open(self.path, "a")
self._wandb = None
if use_wandb:
try:
import wandb
wandb.init(project=wandb_project, name=run_name or f"{stage}-{stamp}", config=config or {})
self._wandb = wandb
except Exception as e: # noqa: BLE001 - logging must never crash training
print(f"[logger] wandb disabled ({e}); JSONL logging only -> {self.path}")
print(f"[logger] stage={stage} -> {self.path}")
def log(self, step: int, metrics: dict[str, Any]) -> None:
record = {"step": step, "wall": time.time(), **metrics}
self._fh.write(json.dumps(record) + "\n")
self._fh.flush()
if self._wandb is not None:
self._wandb.log(metrics, step=step)
def close(self) -> None:
try:
self._fh.close()
finally:
if self._wandb is not None:
self._wandb.finish()