How do you test HE after fine-tuned on CodeLLama?
swtheing opened this issue · comments
swtheing commented
We use 110k data to fine-tune CodeLLama and test the HE and the pass@1 is 60.6%. I wonder some settings are not same between our experiments.
Yuxiang Wei commented
Hi, we finetuned CodeLlama-Python-7B with 75K OSS-Instruct first for 2 epochs and then 110K evol-instruct data for another 2 epochs. The hyperparameter setting is given in the paper and the HumanEval prompt we used is showed at experiments/text2code.py.
Hope this can help with your case and we will later document how to reproduce the experimental results in the paper.