FAQ : Does "InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py" support InternVL3.5 training ?

[HuangChiEn]

I had saw someone applied "InternVL/internvl_chat/internvl/train/internvl_chat_finetune.py" to finetune InternVL3.5 and but i want to confirm is it a official way to do that ?

check list :

1. Conversation template is "internvl2_5", is it valid for InternVL3.5 ?
2. Preprocess function is "internvl2_5", is it valid for InternVL3.5 ?

Any suggestion may be helpful ~

[wuzhaodongaipython]

我也需要微调3.5，我感觉是不是得找sh文件？是哪个sh文件呢，你有找到不

[CrazyElements]

Good questions. I see InternVL3.0's conv_type is also internvl2_5, so i guess conv_type = internvl2_5 is general? And btw, in internvl_chat_gpt_oss/shell there exists scripts for internvl3_5 fine-tuning. Hope this could help you.

[HuangChiEn]

Good questions. I see InternVL3.0's conv_type is also internvl2_5, so i guess conv_type = internvl2_5 is general? And btw, in internvl_chat_gpt_oss/shell there exists scripts for internvl3_5 fine-tuning. Hope this could help you.

我也需要微调3.5，我感觉是不是得找sh文件？是哪个sh文件呢，你有找到不

如同CrazyElements所說，目前我也是用 conv_type = internvl2_5，但微調完後3.5後期爆nan loss，雖然也有可能是我的資料集或multiple loss的影響，至少它應該是可以train ~

至於sh，我是fork其他專案，他們自己有實現資料載入，img token替換等，所以很遺憾我只要找到 conv_type的對應，剩下的整合就在專案內完成，所以我沒有使用到internvl啟動訓練的腳本

[HuangChiEn]

Actually there's one way, but now i'm too lazy to do so ~ Officially, internVL3.5 apply apply_chat_template, which do the same mechanism as Conversation class, so we can inspect the jinja template, and make a equivalent routine in Conversation class will be fine ~