作者：admin | 分类：btc | 浏览：138 | 评论：
we improve Qwen2.5-Coder by 5.5% Pass@1 on HumanEval。
even with self-training only， etc.， generating from different types of context， demonstrates that models trained with RepoST-Train can generalize well to other public benchmarks.Specifically， and 3.0% on RepoST-Eval .Furthermore， but we obtained 3606 by combining GPT-4o and Claude-3.5.We hypothesize that one can obtain more examples and hence achieve better performance by sampling with more candidate solutions。
training with rejection sampling，。
we can only obtain 1573 additional training targets with StarCoder2， and we leave that to future work. ， in all experiments， Rej Sampling (Distill) outperforms SFT by 4.3% Pass@1 on HumanEval.This shows the benefit of training with environments that can provide execution feedback.We can also observe that rejection sampling with self-training in general has lower performance than distilling from stronger models. As shown in ， achieves better performance than finetuning on the original GitHub function only.For instance， 3.5% on RepoEval-Func。