Merge pull request #249 from ModelTC/vlm

docs/en/source/advanced/VLM_quant&img-txt_dataset.md

@@ -60,4 +60,14 @@ Example format of img_qa.json:
 ]
 ```
 The "answer" field is optional.
-The img-txt dataset can include calibration data that contains only text (except for llama3.2).
+The img-txt dataset can include calibration data that contains only text (except for llama3.2).
+
+## VLM Evaluation
+
+LLMC integrates [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval) for evaluations on various downstream datasets. In the config's eval section, the type should be specified as "vqa", and the downstream evaluation datasets in the name should follow the standards set by lmms-eval.
+
+```
+eval:
+    type: vqa
+    name: [mme] # vqav2, gqa, vizwiz_vqa, scienceqa, textvqa
+```

docs/zh_cn/source/advanced/VLM_quant&img-txt_dataset.md

@@ -64,4 +64,14 @@ img_qa.json 格式示例：
 ```
 "answer" 可以不需要
 
-img-txt数据集中可以存在仅有文本的校准数据（当前llama3.2除外）
+img-txt数据集中可以存在仅有文本的校准数据（当前llama3.2除外）
+
+## VLM 测评
+
+llmc接入了[lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval)进行各种下游数据集测评，在config的eval中需要指定type为vqa，name中的下游测评数据集参考lmms-eval的标准。
+
+```
+eval:
+    type: vqa
+    name: [mme] # vqav2, gqa, vizwiz_vqa, scienceqa, textvqa
+```