Li Xiaozhe comments

Results 10 comments of


                                            Li Xiaozhe

[Usage] finetune my data ，Why is the loss not decreasing?

Same question! I also find that on finetine_lora.sh script provided by haotianliu learning rate is 2e-4, so we should change it to 2e-5?

[LLaVA bench in the wild] which result to report?

Same question! Which one to report? There is no 'key | cand/anchor | anchor | cand 'in my output.

[Question] loss 不能下降

Same question

[Question] loss 不能下降

I change learning rate from 2e-4 to 2e-5. It is really work!

[Question] Cannot reproduce lora-mme

Using lora finetuing on llava_mix_665k?

[Question] I finetuning llava1.5-7b model using lora on llava_mix_665k dataset but get 545 on MME perception score. Has anyone else encountered this issue?

@haotian-liu My training hyperparameters remain consistent with you provided. Here are my partial train logs and MME results: 100%|██████████| 10396/10396 [22:39:17

MotionLoRA training

Some question! Could you tell me how to train motionlora?

Do you have the code for the visualization?

Same question！

No image input question

补充：在Qwen2.5VL上面正常，感觉这有可能是Qwen3-VL的image.processor的问题？它处理image后为空，导致它走了text only的分支

No image input question

补充：使用vllm+megatron运行成功，但是sglang+fsdp仍然报错