examples: add RAIL Score responsible AI custom task template

[SumitVermakgp]

Summary

Add a custom task template that evaluates model outputs across 8 responsible AI dimensions using RAIL Score as a MetricGrouping metric.

Changes

• examples/custom_tasks_templates/custom_rail_score_task.py -- Complete self-contained file with RAILScoreComputation(SampleLevelComputation), MetricGrouping registration, prompt function, and task config

How it works

• RAILScoreComputation calls the RAIL Score API on each generated response
• Returns 9 named metrics (8 dimensions + overall), each normalized to 0-1
• Uses MetricGrouping so all dimensions appear as separate columns in results
• Domain context read from doc.specific metadata for domain-specific scoring
• Lazy client initialization (no API key required at import time)

Usage

pip install rail-score-sdk
export RAIL_API_KEY="rail_..."

lighteval accelerate \
    "model_name=HuggingFaceH4/zephyr-7b-beta" \
    "rail_score:default|0" \
    --custom-tasks examples/custom_tasks_templates/custom_rail_score_task.py

Testing

• Tested RAILScoreComputation.compute() with mock Doc and ModelResponse objects
• All 9 metrics returned in correct 0-1 range
• extend_enum registration verified
• Edge cases (short output, empty response) handled
• Formatted with ruff

Related

• Issue: Custom task: RAIL Score responsible AI evaluation (8-dimension scoring) #1202