[MNLI example] Prevent overwriting matched with mismatched metrics

[eldarkurtic]

What does this PR do?

Fixes #16474.

The fix appends _mm to the metrics for mnli-mm evaluation dataset to prevent overwriting eval metrics from the mnli dataset, and writes them together in eval_results.json and all_results.json. The MNLI-metrics look like this:

{
    "eval_accuracy": 0.36067244014263883,
    "eval_accuracy_mm": 0.35506509357200977,
    "eval_loss": 1.158889889717102,
    "eval_loss_mm": 1.1670204401016235,
    "eval_runtime": 16.4691,
    "eval_runtime_mm": 15.9496,
    "eval_samples": 9815,
    "eval_samples_mm": 9832,
    "eval_samples_per_second": 595.964,
    "eval_samples_per_second_mm": 616.44,
    "eval_steps_per_second": 18.641,
    "eval_steps_per_second_mm": 19.311
}

@sgugger, @patil-suraj

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[sgugger]

Thanks a lot for your PR! I just have two comments and it should be good to merged.

Also make sur to run make style on your branch when you're done so the quality check passes.

[sgugger]

I'd regroup this slightly differently:

Suggested change

	if task == "mnli-mm":
	metrics = {k + "_mm": v for k, v in metrics.items()}
	if "mnli" in task:
	combined.update(metrics)
	if task == "mnli-mm":
	combined.update({k + "_mm": v for k, v in metrics.items()})
	elif task == "mnli":
	combined.update(metrics)

[eldarkurtic]

With this regrouping, the trainer.log_metrics("eval", metrics) won't have the "_mm" , so the stdout will be the same as before (that's the main reason why keys in metrics are updated). Given this, should we still regroup and affect only the json log files, but not the stdout?

[sgugger]

Ah understood. LGTM as is then :-)

[sgugger]

The CI failure is unrelated to this PR (actually investigating it right now), so we can merge safely :-)