Gong Zi

We apologize for the delayed response. To address your issue, please try the following: - Use 'qwen' as the `model_type` to load models based on Qwen and Qwen1.5 versions. -...

We apologize for the delayed response. For evaluating the HumanEval metric, as well as a variety of other code-related metrics, we recommend referring to the codefuse-evaluation repository (https://github.com/codefuse-ai/codefuse-evaluation).

We apologize for the delayed response. To address your issue, please follow these steps: 1. First, use the `run_offline_tokenization.sh` script to tokenize your data. 2. Then, make the following modifications...

We apologize for the delayed response. We have provided detailed explanations for each CoBa parameter, along with recommended settings, in the "CoBa Arguments Configuration" section of the mftcoder_accelerate README. These...

> [#2919](https://github.com/volcengine/verl/issues/2919) - Suggested a fix in an issue which I raised. Maybe that would fix your issue Thanks for your reply! I have tried passing `n_micro_batches` as a parameter...

> Hope you removed the division by n micro batches in below line too for step loss logging > > ``` > for micro_batch in micro_batches: > loss = self._compute_loss_and_backward(batch=micro_batch,...

Yup, I noticed this before and tried it, but the loss was still inconsistent. But I didn't try it with performing `loss /= n_micro_batches` before `loss.backward()`. I'll try it again....

But I think we should keep `grad_scaler=True`, because the loss of experiment **sft-qwen3-32b-lr5e-5-32k-gpu64-bsz32-sp2** is correct.

Hi, thanks for the follow-up. After the ablation runs, I’m afraid I haven’t been able to pinpoint the root cause for the inconsistent loss.

This issue may not be related to which MFT loss was used. It's possible that the problem stems from an incorrect setting of the model_type (qwen or qwen2). Could you...