Fix MMLU 5-shot evaluation in evaluator

[joellidin]

Description

The evaluator’s MMLU 5-shot evaluation is currently not working. The problem occurs when inputs are truncated by vLLM due to context length limits, which breaks few-shot evaluation. This prevents us from getting valid MMLU 5-shot metrics for the latest checkpoints.

Tasks

• Investigate how vLLM handles input truncation during few-shot evaluation.
• Adjust evaluator configuration or preprocessing so that MMLU 5-shot prompts fit within the allowed context length.
• Verify that MMLU 5-shot results are produced and logged correctly.

Acceptance Criteria

• Evaluator can successfully run MMLU 5-shot evaluation without truncation errors.
• Correct metrics are reported for MMLU 5-shot in both WandB and InfluxDB.
• Zero-shot and other task evaluations remain unaffected.