Investigate and run lm-eval with few-shot evaluations

[joellidin]

Description

We currently only report zero-shot metrics for checkpoints. To better understand model performance, we should investigate and run lm-evaluation-harness with few-shot settings (e.g., k-shot) on the latest checkpoints. The goal is to report few-shot results on the same benchmark tasks we already track.

Tasks

• Run lm-eval with few-shot configurations on the latest checkpoints.
• Collect and summarize few-shot metrics (accuracy, acc_norm, etc.) for standard tasks (arc, hellaswag, winogrande, …).
• Log results in a consistent way with existing evaluation reports.

Acceptance Criteria

• Few-shot metrics are produced and available for recent checkpoints.
• Results are clearly distinguishable from zero-shot runs.
• Report covers the same benchmark tasks as our existing evaluation flow.