Training Epches for LLM

[andypinxinliu]

Do you know how many training epoches should we do for language model part for different stages? I saw all of them were using 200 epoches, and training is kind of very slow. Do we really need to have so many training epochs for each stage?

[RobinWitch]

No need for it. This is just to ensure convergence during the experimentation process. Generally, once you feel the loss has converged, you can stop it.

[masonwang513]

@RobinWitch May I ask how many epochs you actually trained for the released ckpts, for instance, MECo_BEAT2_2_qwen2.5_0.5b_stage2

[RobinWitch]

@RobinWitch May I ask how many epochs you actually trained for the released ckpts, for instance, MECo_BEAT2_2_qwen2.5_0.5b_stage2

@masonwang513 In my personal training, I achieved the best results at around the 10th epoch. You can save multiple checkpoints, then use the test script to evaluate different checkpoints several times each, and select the one that gives you the best results.

[masonwang513]

@RobinWitch May I ask how many epochs you actually trained for the released ckpts, for instance, MECo_BEAT2_2_qwen2.5_0.5b_stage2

@masonwang513 In my personal training, I achieved the best results at around the 10th epoch. You can save multiple checkpoints, then use the test script to evaluate different checkpoints several times each, and select the one that gives you the best results.

Got it, thanks.