CDIS: Training-Free Data Selection for Instruction Tuning with Compression Ratio and Token Entropy
Submitted to ACL ARR 2026 October on October 5, 2026
CDIS selects instruction-tuning data without additional model training by combining compression ratio and token entropy.
Recommended citation: Jiang, H.*, Wang, Z., Hou, M., Li, X., Guo, T., Zheng, J., & Luo, R. (2026). "CDIS: Training-Free Data Selection for Instruction Tuning with Compression Ratio and Token Entropy." ACL ARR 2026 October. Conference submission. * Equal contribution.