Log per-sample details as trackio.Trace in push_to_wandb

[abidlabs]

Summary

Lighteval already has a Trackio path (import trackio as wandb in evaluation_tracker.py), but two gaps remained that made it less useful than it could be:

1. push_to_wandb accepts details_datasets but never uses it (line 297) — only aggregated results are logged. Per-sample data is dropped on the floor for both wandb and Trackio users.
2. trackio.Trace is unused, even though lighteval's per-sample shape (Doc.query + ModelResponse.text + metric dict + gold_index/choices) maps cleanly onto a conversational trace.

This PR closes both: when Trackio is the active backend, each per-sample detail is emitted as a trackio.Trace(messages=..., metadata=...), so individual eval samples are inspectable on the dashboard alongside the aggregated metrics.

The wandb code path is unchanged.

Trace shape

For each row in a task's details_datasets:

trackio.Trace(
    messages=[
        # only if doc.instruction is set:
        {"role": "system", "content": doc.instruction},
        {"role": "user",      "content": model_response.input or doc.query},
        {"role": "assistant", "content": model_response.text[0]},
    ],
    metadata={
        "task": task_name,
        **metric,                # all per-sample metric values
        "gold_index": doc.gold_index,
        "gold": doc.choices[doc.gold_index],   # when resolvable
    },
)

Logged under {task_name}/sample so samples group by task on the dashboard.

Changes

• src/lighteval/logging/evaluation_tracker.py:
  • Track whether the active backend is Trackio (self._using_trackio) and stash the imported module on the tracker.
  • Add _log_details_as_traces and call it from push_to_wandb behind the trackio check.

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.