Skip to content

[Eval] Customization Bench — domain adaptation evaluation #2450

Description

@sephmard

Summary

Add Customization Bench as a new evaluation in NeMo Gym. This benchmark measures how readily a base or GA model can be fine-tuned (SFT) to a specific domain (e.g., legal, finance) — assessing customizability as a first-class model property alongside raw capability.

Motivation

Frontier model comparisons today focus primarily on out-of-the-box performance. Customization Bench fills a gap by evaluating how efficiently a model adapts to domain-specific tasks with supervised fine-tuning, enabling apples-to-apples comparison of base models on their customizability.

Metadata

Metadata

Assignees

No one assigned

    Type

    No type

    Projects

    Status
    Dev Todo

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions