asadfgglie
/

banban-vision-8b-v1.2.1-zh-DPO

Text Generation

Model card Files Files and versions

asadfgglie/banban-vision-8b-v1.2.1-zh-DPO

將asadfgglie/banban-vision-8b-v1.2-zh對asadfgglie/BanBan_2024-10-17-DPO再次DPO訓練後的版本

目前已知問題是中文能力會直接消失

英文能力反而有靠向這個資料集想要的偏好方向

推測是epoch=4導致的訓練過擬合

我原本想說eval loss與train loss都穩步下降的說

結果我等了10小時卻給我這個

哭阿

Downloads last month: 4

Safetensors

Model size

8.36B params

Tensor type

F16

·

Model tree for asadfgglie/banban-vision-8b-v1.2.1-zh-DPO

Base model

xtuner/llava-llama-3-8b-v1_1-transformers

Finetuned

asadfgglie/banban-vision-8b-v1-pretrain-zh

Finetuned

asadfgglie/banban-vision-8b-v1.2-zh

Finetuned

(1)

this model

Dataset used to train asadfgglie/banban-vision-8b-v1.2.1-zh-DPO