andito HF staff ariG23498 HF staff commited on
Commit
190ae46
1 Parent(s): 6613593

add model card (#2)

Browse files

- add model card (d112364f9360c26972d07217a9fd6b7219ae62e1)


Co-authored-by: Aritra Roy Gosthipaty <[email protected]>

Files changed (1) hide show
  1. README.md +148 -178
README.md CHANGED
@@ -1,202 +1,172 @@
1
  ---
2
  library_name: transformers
3
- tags: []
 
 
 
 
 
 
4
  base_model:
5
  - HuggingFaceTB/SmolLM2-1.7B-Instruct
6
  - google/siglip-so400m-patch14-384
7
  ---
 
8
 
9
- # Model Card for Model ID
10
 
11
- <!-- Provide a quick summary of what the model is/does. -->
 
 
 
12
 
 
13
 
 
 
 
 
 
14
 
15
- ## Model Details
16
 
17
- ### Model Description
18
-
19
- <!-- Provide a longer summary of what this model is. -->
20
-
21
- This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
22
-
23
- - **Developed by:** [More Information Needed]
24
- - **Funded by [optional]:** [More Information Needed]
25
- - **Shared by [optional]:** [More Information Needed]
26
- - **Model type:** [More Information Needed]
27
- - **Language(s) (NLP):** [More Information Needed]
28
- - **License:** [More Information Needed]
29
- - **Finetuned from model [optional]:** [More Information Needed]
30
-
31
- ### Model Sources [optional]
32
-
33
- <!-- Provide the basic links for the model. -->
34
-
35
- - **Repository:** [More Information Needed]
36
- - **Paper [optional]:** [More Information Needed]
37
- - **Demo [optional]:** [More Information Needed]
38
 
39
  ## Uses
40
 
41
- <!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
42
-
43
- ### Direct Use
44
-
45
- <!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
46
-
47
- [More Information Needed]
48
-
49
- ### Downstream Use [optional]
50
-
51
- <!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
52
-
53
- [More Information Needed]
54
-
55
- ### Out-of-Scope Use
56
-
57
- <!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
58
-
59
- [More Information Needed]
60
-
61
- ## Bias, Risks, and Limitations
62
-
63
- <!-- This section is meant to convey both technical and sociotechnical limitations. -->
64
-
65
- [More Information Needed]
66
-
67
- ### Recommendations
68
-
69
- <!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
70
-
71
- Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
72
-
73
- ## How to Get Started with the Model
74
-
75
- Use the code below to get started with the model.
76
-
77
- [More Information Needed]
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
78
 
79
  ## Training Details
80
 
81
  ### Training Data
82
 
83
- <!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
84
-
85
- [More Information Needed]
86
-
87
- ### Training Procedure
88
-
89
- <!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
90
-
91
- #### Preprocessing [optional]
92
-
93
- [More Information Needed]
94
-
95
 
96
- #### Training Hyperparameters
97
-
98
- - **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
99
-
100
- #### Speeds, Sizes, Times [optional]
101
-
102
- <!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
103
-
104
- [More Information Needed]
105
 
106
  ## Evaluation
107
 
108
- <!-- This section describes the evaluation protocols and provides the results. -->
109
-
110
- ### Testing Data, Factors & Metrics
111
-
112
- #### Testing Data
113
-
114
- <!-- This should link to a Dataset Card if possible. -->
115
-
116
- [More Information Needed]
117
-
118
- #### Factors
119
-
120
- <!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
121
-
122
- [More Information Needed]
123
-
124
- #### Metrics
125
-
126
- <!-- These are the evaluation metrics being used, ideally with a description of why. -->
127
-
128
- [More Information Needed]
129
-
130
- ### Results
131
-
132
- [More Information Needed]
133
-
134
- #### Summary
135
-
136
-
137
-
138
- ## Model Examination [optional]
139
-
140
- <!-- Relevant interpretability work for the model goes here -->
141
-
142
- [More Information Needed]
143
-
144
- ## Environmental Impact
145
-
146
- <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
147
-
148
- Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700).
149
-
150
- - **Hardware Type:** [More Information Needed]
151
- - **Hours used:** [More Information Needed]
152
- - **Cloud Provider:** [More Information Needed]
153
- - **Compute Region:** [More Information Needed]
154
- - **Carbon Emitted:** [More Information Needed]
155
-
156
- ## Technical Specifications [optional]
157
-
158
- ### Model Architecture and Objective
159
-
160
- [More Information Needed]
161
-
162
- ### Compute Infrastructure
163
-
164
- [More Information Needed]
165
-
166
- #### Hardware
167
-
168
- [More Information Needed]
169
-
170
- #### Software
171
-
172
- [More Information Needed]
173
-
174
- ## Citation [optional]
175
-
176
- <!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
177
-
178
- **BibTeX:**
179
-
180
- [More Information Needed]
181
-
182
- **APA:**
183
-
184
- [More Information Needed]
185
-
186
- ## Glossary [optional]
187
-
188
- <!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
189
-
190
- [More Information Needed]
191
-
192
- ## More Information [optional]
193
-
194
- [More Information Needed]
195
-
196
- ## Model Card Authors [optional]
197
-
198
- [More Information Needed]
199
-
200
- ## Model Card Contact
201
-
202
- [More Information Needed]
 
1
  ---
2
  library_name: transformers
3
+ license: apache-2.0
4
+ datasets:
5
+ - HuggingFaceM4/the_cauldron
6
+ - HuggingFaceM4/Docmatix
7
+ pipeline_tag: image-text-to-text
8
+ language:
9
+ - en
10
  base_model:
11
  - HuggingFaceTB/SmolLM2-1.7B-Instruct
12
  - google/siglip-so400m-patch14-384
13
  ---
14
+ <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/SmolVLM.png" width="800" height="auto" alt="Image description">
15
 
16
+ # SmolVLM
17
 
18
+ SmolVLM is a compact open multimodal model that accepts arbitrary sequences of image and text inputs to produce text outputs.
19
+ Designed for efficiency, SmolVLM can answer questions about images, describe visual content, create stories grounded on multiple images,
20
+ or function as a pure language model without visual inputs. Its lightweight architecture makes it suitable for on-device applications
21
+ while maintaining strong performance on multimodal tasks.
22
 
23
+ ## Model Summary
24
 
25
+ - **Developed by:** Hugging Face 🤗
26
+ - **Model type:** Multi-modal model (image+text)
27
+ - **Language(s) (NLP):** English
28
+ - **License:** Apache 2.0
29
+ - **Architecture:** Based on [Idefics3](https://huggingface.co/HuggingFaceM4/Idefics3-8B-Llama3) (see technical summary)
30
 
31
+ ## Resources
32
 
33
+ - **Demo:** [SmolVLM Demo](https://huggingface.co/spaces/HuggingFaceTB/SmolVLM)
34
+ - **Blog:** [Blog post](https://huggingface.co/blog/smolvlm)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
35
 
36
  ## Uses
37
 
38
+ SmolVLM can be used for inference on multimodal (image + text) tasks where the input comprises text queries along with one or more images.
39
+ Text and images can be interleaved arbitrarily, enabling tasks like image captioning, visual question answering, and storytelling based on
40
+ visual content. The model does not support image generation.
41
+
42
+ To fine-tune SmolVLM on a specific task, you can follow the fine-tuning tutorial.
43
+ <!-- todo: add link to fine-tuning tutorial -->
44
+
45
+ ### Technical Summary
46
+
47
+ SmolVLM leverages the lightweight SmolLM2 language model to provide a compact yet powerful multimodal experience.
48
+ It introduces several changes compared to previous Idefics models:
49
+
50
+ - **Image compression:** We introduce a more radical image compression compared to Idefics3 to enable the model to infer faster and use less RAM.
51
+ - **Visual Token Encoding:** SmolVLM uses 81 visual tokens to encode image patches of size 384×384. Larger images are divided into patches, each encoded separately, enhancing efficiency without compromising performance.
52
+
53
+ More details about the training and architecture are available in our technical report.
54
+
55
+
56
+ ### How to get started
57
+
58
+ You can use transformers to load, infer and fine-tune SmolVLM.
59
+
60
+ ```python
61
+ import torch
62
+ from PIL import Image
63
+ from transformers import AutoProcessor, AutoModelForVision2Seq
64
+ from transformers.image_utils import load_image
65
+ DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
66
+ # Load images
67
+ image1 = load_image("https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg")
68
+ image2 = load_image("https://huggingface.co/spaces/merve/chameleon-7b/resolve/main/bee.jpg")
69
+ # Initialize processor and model
70
+ processor = AutoProcessor.from_pretrained("HuggingFaceTB/SmolVLM-Synthetic")
71
+ model = AutoModelForVision2Seq.from_pretrained(
72
+ "HuggingFaceTB/SmolVLM-Synthetic",
73
+ torch_dtype=torch.bfloat16,
74
+ _attn_implementation="flash_attention_2" if DEVICE == "cuda" else "eager",
75
+ ).to(DEVICE)
76
+ # Create input messages
77
+ messages = [
78
+ {
79
+ "role": "user",
80
+ "content": [
81
+ {"type": "image"},
82
+ {"type": "image"},
83
+ {"type": "text", "text": "Can you describe the two images?"}
84
+ ]
85
+ },
86
+ ]
87
+ # Prepare inputs
88
+ prompt = processor.apply_chat_template(messages, add_generation_prompt=True)
89
+ inputs = processor(text=prompt, images=[image1, image2], return_tensors="pt")
90
+ inputs = inputs.to(DEVICE)
91
+ # Generate outputs
92
+ generated_ids = model.generate(**inputs, max_new_tokens=500)
93
+ generated_texts = processor.batch_decode(
94
+ generated_ids,
95
+ skip_special_tokens=True,
96
+ )
97
+ print(generated_texts[0])
98
+ """
99
+ User:<image>Can you describe the two images?
100
+ Assistant: The two images are not described in the provided facts, so we cannot provide any information about them.
101
+ """
102
+ ```
103
+
104
+
105
+ ### Model optimizations
106
+
107
+ **Precision**: For better performance, load and run the model in half-precision (`torch.float16` or `torch.bfloat16`) if your hardware supports it.
108
+
109
+ ```python
110
+ from transformers import AutoModelForVision2Seq
111
+ import torch
112
+ model = AutoModelForVision2Seq.from_pretrained(
113
+ "HuggingFaceTB/SmolVLM-Synthetic",
114
+ torch_dtype=torch.bfloat16
115
+ ).to("cuda")
116
+ ```
117
+
118
+ You can also load SmolVLM with 4/8-bit quantization using bitsandbytes, torchao or Quanto. Refer to [this page](https://huggingface.co/docs/transformers/en/main_classes/quantization) for other options.
119
+
120
+ ```python
121
+ from transformers import AutoModelForVision2Seq, BitsAndBytesConfig
122
+ import torch
123
+ quantization_config = BitsAndBytesConfig(load_in_8bit=True)
124
+ model = AutoModelForVision2Seq.from_pretrained(
125
+ "HuggingFaceTB/SmolVLM-Synthetic",
126
+ quantization_config=quantization_config,
127
+ )
128
+ ```
129
+
130
+ **Vision Encoder Efficiency**: Adjust the image resolution by setting `size={"longest_edge": N*384}` when initializing the processor, where N is your desired value. The default `N=4` works well, which results in input images of
131
+ size 1536×1536. For documents, `N=5` might be beneficial. Decreasing N can save GPU memory and is appropriate for lower-resolution images. This is also useful if you want to fine-tune on videos.
132
+
133
+
134
+ ## Misuse and Out-of-scope Use
135
+
136
+ SmolVLM is not intended for high-stakes scenarios or critical decision-making processes that affect an individual's well-being or livelihood. The model may produce content that appears factual but may not be accurate. Misuse includes, but is not limited to:
137
+
138
+ - Prohibited Uses:
139
+ - Evaluating or scoring individuals (e.g., in employment, education, credit)
140
+ - Critical automated decision-making
141
+ - Generating unreliable factual content
142
+ - Malicious Activities:
143
+ - Spam generation
144
+ - Disinformation campaigns
145
+ - Harassment or abuse
146
+ - Unauthorized surveillance
147
+
148
+ ### License
149
+
150
+ SmolVLM is built upon [the shape-optimized SigLIP](https://huggingface.co/google/siglip-so400m-patch14-384) as image encoder and [SmolLM2](https://huggingface.co/HuggingFaceTB/SmolLM2-1.7B-Instruct) for text decoder part.
151
+
152
+ We release the SmolVLM checkpoints under the Apache 2.0 license.
153
 
154
  ## Training Details
155
 
156
  ### Training Data
157
 
158
+ The training data comes from [The Cauldron](https://huggingface.co/datasets/HuggingFaceM4/the_cauldron) and [Docmatix](https://huggingface.co/datasets/HuggingFaceM4/Docmatix) datasets, with emphasis on document understanding (25%) and image captioning (18%), while maintaining balanced coverage across other crucial capabilities like visual reasoning, chart comprehension, and general instruction following.
159
+ <img src="https://huggingface.co/HuggingFaceTB/SmolVLM-Instruct/resolve/main/mixture_the_cauldron.png" alt="Example Image" style="width:90%;" />
 
 
 
 
 
 
 
 
 
 
160
 
 
 
 
 
 
 
 
 
 
161
 
162
  ## Evaluation
163
 
164
+ | Model | MMMU (val) | MathVista (testmini) | MMStar (val) | DocVQA (test) | TextVQA (val) | Min GPU RAM required (GB) |
165
+ |-------------------|------------|----------------------|--------------|---------------|---------------|---------------------------|
166
+ | SmolVLM | 38.8 | 44.6 | 42.1 | 81.6 | 72.7 | 5.02 |
167
+ | Qwen-VL 2B | 41.1 | 47.8 | 47.5 | 90.1 | 79.7 | 13.70 |
168
+ | InternVL2 2B | 34.3 | 46.3 | 49.8 | 86.9 | 73.4 | 10.52 |
169
+ | PaliGemma 3B 448px| 34.9 | 28.7 | 48.3 | 32.2 | 56.0 | 6.72 |
170
+ | moondream2 | 32.4 | 24.3 | 40.3 | 70.5 | 65.2 | 3.87 |
171
+ | MiniCPM-V-2 | 38.2 | 39.8 | 39.1 | 71.9 | 74.1 | 7.88 |
172
+ | MM1.5 1B | 35.8 | 37.2 | 0.0 | 81.0 | 72.5 | NaN |