MT-Bench-101 (1/2): Declare MT-Bench-101 task and dataset adapter

[ErlisLushtaku]

Summary

This is the first PR in the MT-Bench-101 stack. It declares the task and adds the dataset adapter; generation and judging are in #120.

Changes

• Add a packaged mt-bench-101 task YAML with the upstream dataset pinned to commit bc18b3e2c18c99164e11528f1a79c92083db5953.
• Expand the source dialogues into one row per evaluated turn while retaining the golden context. Tasks that only become meaningful after the first turn start at turn 2, following the reference implementation.
• Add the typed protocol schema and register the dataset adapter and runner route.
• Package the task-specific judge prompts used by the OpenCompass fork, with the typo and formatting corrections carried over from closed feat: add support for mt-bench-101 #22.

References

The task structure and prompts come from the MT-Bench-101 repository, the OpenCompass fork, and Bai et al. (2024).

[kargibora]

What is GoldenContext?

[ErlisLushtaku]

It means each turn is generated using the preceding reference dialogue as fixed context, rather than using the previously generated turns from the LLM. I added this to the schema.

[kargibora]

Looks good to me as it is just the definitions. IT is a clean implementation!

[kargibora]

Thanks for clarifiyng!