codeql
qwen3-4b-refiner-codeql-self-nothink-fullqwen3-8b-refiner-codeql-self-nothink-fullqwen3-0.6b-refiner-codeql-self-nothink-fullqwen3-8b-grpo-refiner-codeql-exec-global_step_850qwen3-0.6b-refiner-codeql-self-nothink-full-Q8_0-GGUFqwen3-4b-grpo-refiner-codeql-exec-global_step_850qwen3-0.6b-grpo-refiner-codeql-exec-global_step_932qwen3-0.6b-grpo-refiner-codeql-exec-global_step_800
Datasets
All datasets matching “codeql”CSCAN-CodeQL
CSCAN-CodeQL
CodeQL vulnerability detection
Attribution
Author: Euisuh JeongAffiliation: Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa UniversityLicense: MIT
Citation
@dataset{cscan_codeql,
author={Jeong, Euisuh},
year={2026},
title={CSCAN-CodeQL},
publisher={Hugging Face},
howpublished={\url{https://huggingface.co/datasets/euisuh/CSCAN-CodeQL}}
}
codeql1_aggressivecodeql2_aggressiveverl_codeql1_aggressive_filtered
verl_codeql1_aggressive_filtered
Curriculum-filtered subset of OctoReasoner/verl_codeql1_aggressive.
CoT variant. Models were prompted to think first, then box (max_tokens=4096). Kept rows satisfy weak <= 5/8 AND strong >= 2/8, where weak/strong = number of correct rollouts (of 8, strict \boxed{} scoring) from Qwen2.5-Coder-1.5B/-32B-Instruct.
5591 rows kept. Pure verl schema.
