Run1 skill 1
[COLM'26] SkillLearnBench is the first benchmark for evaluating continual learning methods that automatically generate agent skills.From the repository description
npx -y skills add cxcscmu/SkillLearnBench --skill run1_skill-1Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.
SKILL.md
0.8 KB, 154 tokens by cl100k_base, as published. Nobody here has run it
[SKILL]
name: simpo-preference-optimization description: Detailed explanation and implementation guide for the SimPO (Simple Preference Optimization) loss function, which uses length-normalized log probabilities and a reference-free reward margin.
SimPO (Simple Preference Optimization)
SimPO is a reference-free preference optimization algorithm that simplifies alignment compared to methods like DPO (Direct Preference Optimization). It eliminates the need for a reference model by defining the reward directly using the policy model's length-normalized log probabilities.
Mathematical Formulation
In SimPO, the reward $r_\theta(y)$ for a generated response $y$ given a prompt $x$ is defined as the average log probability of the response tokens, scaled by a constant $\beta$:
What ships with it
Read from the repository
Just SKILL.md. No reference files, no scripts.