Datasets:
tianyu-z
commited on
Commit
•
6e3e253
1
Parent(s):
a36bbc3
README.md
CHANGED
@@ -65,12 +65,14 @@ EM means `"Exact Match"` and Jaccard means `"Jaccard Similarity"`. The best in c
|
|
65 |
| GPT-4 Turbo | - | *78.74* | *88.54* | *45.15* | *65.72* | 0.2 | 8.42 | 0.0 | *8.58* |
|
66 |
| GPT-4V | - | 52.04 | 65.36 | 25.83 | 44.63 | - | - | - | - |
|
67 |
| GPT-4o | - | **91.55** | **96.44** | **73.2** | **86.17** | **14.87** | **39.05** | **2.2** | **22.72** |
|
|
|
68 |
| Gemini 1.5 Pro | - | 62.73 | 77.71 | 28.07 | 51.9 | 1.1 | 11.1 | 0.7 | 11.82 |
|
69 |
| Qwen-VL-Max | - | 76.8 | 85.71 | 41.65 | 61.18 | *6.34* | *13.45* | *0.89* | 5.4 |
|
70 |
| Reka Core | - | 66.46 | 84.23 | 6.71 | 25.84 | 0.0 | 3.43 | 0.0 | 3.35 |
|
71 |
| Cambrian-1 | 34B | 79.69 | 89.27 | *27.20* | 50.04 | 0.03 | 1.27 | 0.00 | 1.37 |
|
72 |
| Cambrian-1 | 13B | 49.35 | 65.11 | 8.37 | 29.12 | - | - | - | - |
|
73 |
| Cambrian-1 | 8B | 71.13 | 83.68 | 13.78 | 35.78 | - | - | - | - |
|
|
|
74 |
| CogVLM2 | 19B | *83.25* | *89.75* | **37.98** | **59.99** | 9.15 | 17.12 | 0.08 | 3.67 |
|
75 |
| CogVLM2-Chinese | 19B | 79.90 | 87.42 | 25.13 | 48.76 | **33.24** | **57.57** | **1.34** | **17.35** |
|
76 |
| DeepSeek-VL | 1.3B | 23.04 | 46.84 | 0.16 | 11.89 | 0.0 | 6.56 | 0.0 | 6.46 |
|
@@ -81,9 +83,11 @@ EM means `"Exact Match"` and Jaccard means `"Jaccard Similarity"`. The best in c
|
|
81 |
| InternLM-XComposer2-VL | 7B | 46.64 | 70.99 | 0.7 | 12.51 | 0.27 | 12.32 | 0.07 | 8.97 |
|
82 |
| InternLM-XComposer2-VL-4KHD | 7B | 5.32 | 22.14 | 0.21 | 9.52 | 0.46 | 12.31 | 0.05 | 7.67 |
|
83 |
| InternLM-XComposer2.5-VL | 7B | 41.35 | 63.04 | 0.93 | 13.82 | 0.46 | 12.97 | 0.11 | 10.95 |
|
84 |
-
| InternVL-V1.5 |
|
85 |
| InternVL-V2 | 26B | 74.51 | 86.74 | 6.18 | 24.52 | 9.02 | 32.50 | 0.05 | 9.49 |
|
86 |
| InternVL-V2 | 40B | **84.67** | **92.64** | 13.10 | 33.64 | 22.09 | 47.62 | 0.48 | 12.57 |
|
|
|
|
|
87 |
| MiniCPM-V2.5 | 8B | 31.81 | 53.24 | 1.41 | 11.94 | 4.1 | 18.03 | 0.09 | 7.39 |
|
88 |
| Monkey | 7B | 50.66 | 67.6 | 1.96 | 14.02 | 0.62 | 8.34 | 0.12 | 6.36 |
|
89 |
| Qwen-VL | 7B | 49.71 | 69.94 | 2.0 | 15.04 | 0.04 | 1.5 | 0.01 | 1.17 |
|
@@ -92,39 +96,43 @@ EM means `"Exact Match"` and Jaccard means `"Jaccard Similarity"`. The best in c
|
|
92 |
|
93 |
# Model Evaluation
|
94 |
|
95 |
-
## Method 1
|
96 |
-
```bash
|
97 |
-
git clone https://github.com/tianyu-z/VCR.git
|
98 |
-
```
|
99 |
### Open-source evaluation
|
100 |
We support open-source model_id:
|
101 |
```python
|
102 |
["openbmb/MiniCPM-Llama3-V-2_5",
|
103 |
"OpenGVLab/InternVL-Chat-V1-5",
|
104 |
"internlm/internlm-xcomposer2-vl-7b",
|
|
|
|
|
105 |
"HuggingFaceM4/idefics2-8b",
|
106 |
"Qwen/Qwen-VL-Chat",
|
107 |
"THUDM/cogvlm2-llama3-chinese-chat-19B",
|
108 |
"THUDM/cogvlm2-llama3-chat-19B",
|
109 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
110 |
```
|
111 |
For the models not on list, they are not intergated with huggingface, please refer to their github repo to create the evaluation pipeline. Examples of the inference logic are in `src/evaluation/inference.py`
|
112 |
|
113 |
```bash
|
114 |
pip install -r requirements.txt
|
115 |
# We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
|
116 |
-
# Inference from the VLMs and save the results to {model_id}_{difficulty}_{language}.json
|
117 |
cd src/evaluation
|
118 |
-
python3 inference.py --dataset_handler "vcr-org/VCR-wiki-en-easy-test" --model_id "HuggingFaceM4/idefics2-8b" --device "cuda" --dtype "bf16" --save_interval 50 --resume True
|
119 |
-
|
120 |
# Evaluate the results and save the evaluation metrics to {model_id}_{difficulty}_{language}_evaluation_result.json
|
121 |
-
python3
|
122 |
-
|
123 |
-
# To get the mean score of all the `{model_id}_{difficulty}_{language}_evaluation_result.json` in `jsons_path` (and the std, confidence interval if `--bootstrap`) of the evaluation metrics
|
124 |
-
python3 gather_results.py --jsons_path .
|
125 |
```
|
|
|
|
|
126 |
|
127 |
-
### Close-source evaluation
|
128 |
We provide the evaluation script for the close-source models in `src/evaluation/closed_source_eval.py`.
|
129 |
|
130 |
You need an API Key, a pre-saved testing dataset and specify the path of the data saving the paper
|
@@ -146,14 +154,46 @@ python3 evaluation_metrics.py --model_id gpt4o --output_path . --json_filename "
|
|
146 |
# To get the mean score of all the `{model_id}_{difficulty}_{language}_evaluation_result.json` in `jsons_path` (and the std, confidence interval if `--bootstrap`) of the evaluation metrics
|
147 |
python3 gather_results.py --jsons_path .
|
148 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
|
150 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
You may need to incorporate the inference method of your model if the lmms-eval framework does not support it. For details, please refer to [here](https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/docs/model_guide.md)
|
152 |
```bash
|
153 |
pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
|
154 |
# We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
|
155 |
python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model idefics2 --model_args pretrained="HuggingFaceM4/idefics2-8b" --tasks vcr_wiki_en_easy --batch_size 1 --log_samples --log_samples_suffix HuggingFaceM4_idefics2-8b_vcr_wiki_en_easy --output_path ./logs/
|
156 |
```
|
|
|
|
|
157 |
`lmms-eval` supports the following VCR `--tasks` settings:
|
158 |
|
159 |
* English
|
|
|
65 |
| GPT-4 Turbo | - | *78.74* | *88.54* | *45.15* | *65.72* | 0.2 | 8.42 | 0.0 | *8.58* |
|
66 |
| GPT-4V | - | 52.04 | 65.36 | 25.83 | 44.63 | - | - | - | - |
|
67 |
| GPT-4o | - | **91.55** | **96.44** | **73.2** | **86.17** | **14.87** | **39.05** | **2.2** | **22.72** |
|
68 |
+
| GPT-4o-mini | - | 83.60 | 87.77 | 54.04 | 73.09 | 1.10 | 5.03 | 0 | 2.02 |
|
69 |
| Gemini 1.5 Pro | - | 62.73 | 77.71 | 28.07 | 51.9 | 1.1 | 11.1 | 0.7 | 11.82 |
|
70 |
| Qwen-VL-Max | - | 76.8 | 85.71 | 41.65 | 61.18 | *6.34* | *13.45* | *0.89* | 5.4 |
|
71 |
| Reka Core | - | 66.46 | 84.23 | 6.71 | 25.84 | 0.0 | 3.43 | 0.0 | 3.35 |
|
72 |
| Cambrian-1 | 34B | 79.69 | 89.27 | *27.20* | 50.04 | 0.03 | 1.27 | 0.00 | 1.37 |
|
73 |
| Cambrian-1 | 13B | 49.35 | 65.11 | 8.37 | 29.12 | - | - | - | - |
|
74 |
| Cambrian-1 | 8B | 71.13 | 83.68 | 13.78 | 35.78 | - | - | - | - |
|
75 |
+
| CogVLM | 17B | 73.88 | 86.24 | 34.58 | 57.17 | - | - | - | - |
|
76 |
| CogVLM2 | 19B | *83.25* | *89.75* | **37.98** | **59.99** | 9.15 | 17.12 | 0.08 | 3.67 |
|
77 |
| CogVLM2-Chinese | 19B | 79.90 | 87.42 | 25.13 | 48.76 | **33.24** | **57.57** | **1.34** | **17.35** |
|
78 |
| DeepSeek-VL | 1.3B | 23.04 | 46.84 | 0.16 | 11.89 | 0.0 | 6.56 | 0.0 | 6.46 |
|
|
|
83 |
| InternLM-XComposer2-VL | 7B | 46.64 | 70.99 | 0.7 | 12.51 | 0.27 | 12.32 | 0.07 | 8.97 |
|
84 |
| InternLM-XComposer2-VL-4KHD | 7B | 5.32 | 22.14 | 0.21 | 9.52 | 0.46 | 12.31 | 0.05 | 7.67 |
|
85 |
| InternLM-XComposer2.5-VL | 7B | 41.35 | 63.04 | 0.93 | 13.82 | 0.46 | 12.97 | 0.11 | 10.95 |
|
86 |
+
| InternVL-V1.5 | 26B | 14.65 | 51.42 | 1.99 | 16.73 | 4.78 | 26.43 | 0.03 | 8.46 |
|
87 |
| InternVL-V2 | 26B | 74.51 | 86.74 | 6.18 | 24.52 | 9.02 | 32.50 | 0.05 | 9.49 |
|
88 |
| InternVL-V2 | 40B | **84.67** | **92.64** | 13.10 | 33.64 | 22.09 | 47.62 | 0.48 | 12.57 |
|
89 |
+
| InternVL-V2 | 76B | 83.20 | 91.26 | 18.45 | 41.16 | 20.58 | 44.59 | 0.56 | 15.31 |
|
90 |
+
| InternVL-V2-Pro | - | 77.41 | 86.59 | 12.94 | 35.01 | 19.58 | 43.98 | 0.84 | 13.97 |
|
91 |
| MiniCPM-V2.5 | 8B | 31.81 | 53.24 | 1.41 | 11.94 | 4.1 | 18.03 | 0.09 | 7.39 |
|
92 |
| Monkey | 7B | 50.66 | 67.6 | 1.96 | 14.02 | 0.62 | 8.34 | 0.12 | 6.36 |
|
93 |
| Qwen-VL | 7B | 49.71 | 69.94 | 2.0 | 15.04 | 0.04 | 1.5 | 0.01 | 1.17 |
|
|
|
96 |
|
97 |
# Model Evaluation
|
98 |
|
99 |
+
## Method 1: use the evaluation script
|
|
|
|
|
|
|
100 |
### Open-source evaluation
|
101 |
We support open-source model_id:
|
102 |
```python
|
103 |
["openbmb/MiniCPM-Llama3-V-2_5",
|
104 |
"OpenGVLab/InternVL-Chat-V1-5",
|
105 |
"internlm/internlm-xcomposer2-vl-7b",
|
106 |
+
"internlm/internlm-xcomposer2-4khd-7b",
|
107 |
+
"internlm/internlm-xcomposer2d5-7b",
|
108 |
"HuggingFaceM4/idefics2-8b",
|
109 |
"Qwen/Qwen-VL-Chat",
|
110 |
"THUDM/cogvlm2-llama3-chinese-chat-19B",
|
111 |
"THUDM/cogvlm2-llama3-chat-19B",
|
112 |
+
"THUDM/cogvlm-chat-hf",
|
113 |
+
"echo840/Monkey-Chat",
|
114 |
+
"THUDM/glm-4v-9b",
|
115 |
+
"nyu-visionx/cambrian-phi3-3b",
|
116 |
+
"nyu-visionx/cambrian-8b",
|
117 |
+
"nyu-visionx/cambrian-13b",
|
118 |
+
"nyu-visionx/cambrian-34b",
|
119 |
+
"OpenGVLab/InternVL2-26B",
|
120 |
+
"OpenGVLab/InternVL2-40B"
|
121 |
+
"OpenGVLab/InternVL2-Llama3-76B",]
|
122 |
```
|
123 |
For the models not on list, they are not intergated with huggingface, please refer to their github repo to create the evaluation pipeline. Examples of the inference logic are in `src/evaluation/inference.py`
|
124 |
|
125 |
```bash
|
126 |
pip install -r requirements.txt
|
127 |
# We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
|
|
|
128 |
cd src/evaluation
|
|
|
|
|
129 |
# Evaluate the results and save the evaluation metrics to {model_id}_{difficulty}_{language}_evaluation_result.json
|
130 |
+
python3 evaluation_pipeline.py --dataset_handler "vcr-org/VCR-wiki-en-easy-test" --model_id HuggingFaceM4/idefics2-8b --device "cuda" --output_path . --bootstrap --end_index 5000
|
|
|
|
|
|
|
131 |
```
|
132 |
+
For large models like "OpenGVLab/InternVL2-Llama3-76B", you may have to use multi-GPU to do the evaluation. You can specify --device to None to use all GPUs available.
|
133 |
+
|
134 |
|
135 |
+
### Close-source evaluation (using API)
|
136 |
We provide the evaluation script for the close-source models in `src/evaluation/closed_source_eval.py`.
|
137 |
|
138 |
You need an API Key, a pre-saved testing dataset and specify the path of the data saving the paper
|
|
|
154 |
# To get the mean score of all the `{model_id}_{difficulty}_{language}_evaluation_result.json` in `jsons_path` (and the std, confidence interval if `--bootstrap`) of the evaluation metrics
|
155 |
python3 gather_results.py --jsons_path .
|
156 |
```
|
157 |
+
## Method 2: use [VLMEvalKit](https://github.com/open-compass/VLMEvalKit) framework
|
158 |
+
You may need to incorporate the inference method of your model if the VLMEvalKit framework does not support it. For details, please refer to [here](https://github.com/open-compass/VLMEvalKit/blob/main/docs/en/Development.md)
|
159 |
+
```bash
|
160 |
+
git clone https://github.com/open-compass/VLMEvalKit.git
|
161 |
+
cd VLMEvalKit
|
162 |
+
# We use HuggingFaceM4/idefics2-8b and VCR_EN_EASY_ALL as an example
|
163 |
+
python run.py --data VCR_EN_EASY_ALL --model idefics2_8b --verbose
|
164 |
+
```
|
165 |
+
You may find the supported model list [here](https://github.com/open-compass/VLMEvalKit/blob/main/vlmeval/config.py).
|
166 |
|
167 |
+
`VLMEvalKit` supports the following VCR `--data` settings:
|
168 |
+
|
169 |
+
* English
|
170 |
+
* Easy
|
171 |
+
* `VCR_EN_EASY_ALL` (full test set, 5000 instances)
|
172 |
+
* `VCR_EN_EASY_500` (first 500 instances in the VCR_EN_EASY_ALL setting)
|
173 |
+
* `VCR_EN_EASY_100` (first 100 instances in the VCR_EN_EASY_ALL setting)
|
174 |
+
* Hard
|
175 |
+
* `VCR_EN_HARD_ALL` (full test set, 5000 instances)
|
176 |
+
* `VCR_EN_HARD_500` (first 500 instances in the VCR_EN_HARD_ALL setting)
|
177 |
+
* `VCR_EN_HARD_100` (first 100 instances in the VCR_EN_HARD_ALL setting)
|
178 |
+
* Chinese
|
179 |
+
* Easy
|
180 |
+
* `VCR_ZH_EASY_ALL` (full test set, 5000 instances)
|
181 |
+
* `VCR_ZH_EASY_500` (first 500 instances in the VCR_ZH_EASY_ALL setting)
|
182 |
+
* `VCR_ZH_EASY_100` (first 100 instances in the VCR_ZH_EASY_ALL setting)
|
183 |
+
* Hard
|
184 |
+
* `VCR_ZH_HARD_ALL` (full test set, 5000 instances)
|
185 |
+
* `VCR_ZH_HARD_500` (first 500 instances in the VCR_ZH_HARD_ALL setting)
|
186 |
+
* `VCR_ZH_HARD_100` (first 100 instances in the VCR_ZH_HARD_ALL setting)
|
187 |
+
|
188 |
+
## Method 3: use lmms-eval framework
|
189 |
You may need to incorporate the inference method of your model if the lmms-eval framework does not support it. For details, please refer to [here](https://github.com/EvolvingLMMs-Lab/lmms-eval/blob/main/docs/model_guide.md)
|
190 |
```bash
|
191 |
pip install git+https://github.com/EvolvingLMMs-Lab/lmms-eval.git
|
192 |
# We use HuggingFaceM4/idefics2-8b and vcr_wiki_en_easy as an example
|
193 |
python3 -m accelerate.commands.launch --num_processes=8 -m lmms_eval --model idefics2 --model_args pretrained="HuggingFaceM4/idefics2-8b" --tasks vcr_wiki_en_easy --batch_size 1 --log_samples --log_samples_suffix HuggingFaceM4_idefics2-8b_vcr_wiki_en_easy --output_path ./logs/
|
194 |
```
|
195 |
+
You may find the supported model list [here](https://github.com/EvolvingLMMs-Lab/lmms-eval/tree/main/lmms_eval/models).
|
196 |
+
|
197 |
`lmms-eval` supports the following VCR `--tasks` settings:
|
198 |
|
199 |
* English
|