Feature request

Currently, the GRPOConfig docstring for log_completions is:

log_completions (bool, optional, defaults to False) — Whether to log the completions during training.

Which is a bit misleading, given that these are only logged when Weights and Biases is active.

Suggestion 1: Change the docs:

log_completions (bool, optional, defaults to False) — If True, will try to log generated (prompt, completion) pairs to wandb, if it's available.

Suggestion 2: Log to wandb if possible, but also log to console. Since rich is a requirement, it could be a rich table:

from rich.console import Console
from rich.panel import Panel
from rich.table import Table
from rich.text import Text

def print_output_sample(prompts: list[str], completions: list[str]):
    """Print out a sample of model completions."""
    console = Console()
    table = Table(show_header=True, header_style="bold white", expand=True, padding=(0, 1, 1, 0))
    table.add_column("Prompt", style="bright_yellow")
    for s, p in zip(source, prediction, strict=True):
        table.add_row(Text(s), Text(p))
    panel = Panel(table, expand=False, title="Output Sample", border_style="bold white")
    console.print(panel)

In that case, it may be useful to let log_completions be an int to specify a custom interval and avoid the console being spammed.

Motivation

It's only right

Your contribution

See suggestions