deepgpt at SemEval-2026 Task 1: A Chinese Humor Generation System via Instruction-Masked QLoRA and Reverse Constraint Data Mixing

AbstractAbstractThis paper presents the system description of the deepgpt team for SemEval2026 Task 1 (MWAHAHA: ComputationalHumor Generation), Subtask A. To address the challenge of generating highquality Chinese humor under strict textconstraints (e.g., incorporating speciffedrare words or relating to news headlines),we propose a parameter-effffcient generation system based on Qwen2.5-3B-Instruct.We reconstructed 8,000 multi-source Chinese jokes into a conversational instruction tuning format. Crucially, to mitigate the prevalent issues of formatting hallucinations and template collapse, we introduce a strict Instruction Masking strategy during 4-bit QLoRA ffne-tuning. Bycompletely isolating the loss calculationto the target humorous text, the modelis forced to treat constraints as conditional inputs rather than conversationaldistributions to mimic. Empirical resultsshow that this architectural interventioncompletely eradicates meaningless conversational ffllers. Our system signiffcantlyboosted the hard constraint adherence (CAcc) to 94.6% and achieved a highly competitive Elo rating of 903 in the offffcialPairwise Human Evaluation, validating theeffectiveness of speciffc masking ffne-tuningfor lightweight large language models instrictly constrained generation tasks.