About the deep-coder-1.5b-base-ckpt

[ck853178967]

I'm trying to reproduce the training process of deepcoder 1.5b, but I found the base model is "checkpoints/deepscaler/deepscaler-code-32k-easy/global_step_320/actor/checkpoint" at https://wandb.ai/mluo/deepcoder/runs/s3lpnxwa/overview, with already relatively high val score (~ 23 on test_lcb)
how to improve from r1-distilled-qwen-1.5b (~ 16-17) to the performance of this ckpt? Thanks for your advices~

[michaelzhiluo]

@ck853178967 Yes, we lost some logs from a prior run.

In paritcular, our starting model was fromDeepseek-1.5b-Qwen -> DeepScaler-1.5B-Preview -> deepscaler-code-32k-easy/global_step_320/actor/checkpoint -> DeepCoder-1.5B-Preview