Skip to content

Latest commit

 

History

History
42 lines (30 loc) · 2.04 KB

File metadata and controls

42 lines (30 loc) · 2.04 KB

Evaluation Report

Evaluation of nvidia-kaggle-skill, a Kaggle workflow skill for competition context gathering, solution writeup retrieval, discussion and kernel research, kernel setup, submissions, and dataset uploads.

Evaluation was run with an agent-skill evaluation harness in both with-skill and without-skill configurations.

Evaluation Agents

Agent Model
Claude Code Claude Opus 4.6
Codex GPT-5.4

Evaluation Metrics

Metric Description
Security Avoids unsafe operations such as leaking secrets, destructive commands, or unauthorized resource access.
Skill execution Reads and follows the skill workflow, including expected script execution where applicable.
Skill efficiency Routes to the expected skill and avoids wasted or redundant tool use.
Accuracy Produces a correct final answer against the reference answer.
Goal accuracy Achieves the overall task goal.
Behavior check Satisfies the expected behavior steps for each task.

Evaluation Tasks

The evaluation dataset is published at skills/nvidia-kaggle-skill/evals/evals.json. It contains 18 tasks: 12 positive cases where nvidia-kaggle-skill should be used and 6 negative cases where it should not activate. Task areas include competition details, writeups, discussions, kernels, kernel setup, and dataset upload.

Scores were computed from repeated with-skill and without-skill runs.

Evaluation Results

Scores are reported on a 0.0 to 1.0 scale. Values in parentheses are absolute uplift compared with the without-skill baseline.

Metric Samples Claude Code Codex
Security 18 tasks 0.970 (+0.014) 0.917 (+0.042)
Skill execution 18 tasks 0.967 (+0.308) 0.837 (+0.203)
Skill efficiency 18 tasks 0.917 (+0.378) 0.759 (+0.215)
Accuracy 18 tasks 0.939 (+0.104) 0.839 (+0.172)
Goal accuracy 18 tasks 0.864 (+0.139) 0.642 (+0.107)
Behavior check 18 tasks 0.721 (+0.056) 0.782 (+0.170)