fine-tuned qwen 2.5 7b on coding tasks, now it refuses to generate any code over 50 lines

spent last week fine-tuning qwen 2.5 7b q4_k_m on a dataset of ~2800 python functions for internal tooling. training went fine, loss curve looked good, no nans, finished all 3 epochs clean. but now the fine-tuned model refuses to generate any code longer than about 50 lines. if you ask for something that would naturally be 100+ lines, it generates the first 40-50 lines and then just stops mid-function with no closing braces or explanation. base model had no issues generating long code. training config: qlora rank 32, lr 1e-4, batch size 4, gradient accumulation 8, max sequence length 2048. dataset examples ranged from 20 to 300 lines of code. i'm assuming this is either (a) the model learned that most training examples were under 100 lines so it overfitted to that distribution, or (b) something about the qlora setup is causing early stopping during generation. but honestly just guessing at this point. has anyone hit this where fine-tuning makes the model worse at generating long outputs even though base was fine at it?