Commit
•
a9d32da
1
Parent(s):
ba821b7
Updating Model card
Browse filesUpdating model card to fit recommended standards. New model card is composed of the content available in the original model card for this repository.
README.md
CHANGED
@@ -2,58 +2,78 @@
|
|
2 |
license: other
|
3 |
license_name: intel-research-use-license
|
4 |
license_link: LICENSE
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
---
|
6 |
|
7 |
-
|
8 |
|
9 |
-
|
10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
11 |
|
12 |
-
|
13 |
|
14 |
-
|
15 |
|
16 |
-
|
|
|
|
|
|
|
|
|
17 |
|
18 |
-
The model has been finetuned for multimodal benchmark evaluations, but can also be used as a multimodal chatbot.
|
19 |
|
20 |
-
|
21 |
|
22 |
-
|
23 |
-
|
24 |
-
## Training Details
|
25 |
-
|
26 |
-
The `llava-llama-3-8b` model was trained on a 4 node cluster with a total of 32 Gaudi 2 accelerators.
|
27 |
-
|
28 |
-
### Training Data
|
29 |
-
|
30 |
-
The model was trained using the LLaVA-v1.5 data mixture.
|
31 |
-
|
32 |
-
This is listed as follows:
|
33 |
-
|
34 |
-
- 558K filtered image-text pairs from LAION/CC/SBU, captioned by BLIP.
|
35 |
-
- 158K GPT-generated multimodal instruction-following data.
|
36 |
-
- 450K academic-task-oriented VQA data mixture.
|
37 |
-
- 40K ShareGPT data.
|
38 |
-
|
39 |
-
## Evaluation
|
40 |
-
|
41 |
-
| Model | Metrics |
|
42 |
-
|----------|------------------|
|
43 |
-
| ScienceQA| 72.9797 |
|
44 |
-
| MMVet | 31.9725 |
|
45 |
-
| llavaw | 56.9/61.9/73.6/65.7 |
|
46 |
-
| Pope Acc | 87.33, F1 86.5 |
|
47 |
-
| GQA | 60.6138 |
|
48 |
-
| MMVP | 36 |
|
49 |
-
|
50 |
-
## License
|
51 |
-
The weights are released under the Intel Research Use License Agreement (see LICENSE file)
|
52 |
-
All usage code is licensed Apache 2.0
|
53 |
-
|
54 |
-
## Usage
|
55 |
-
|
56 |
-
Please note, we only provide the trained weights difference and do not provide a copy of the base [`meta-llama/Meta-Llama-3-8B-Instruct`](https://huggingface.co/meta-llama/Meta-Llama-3-8B) model. Any use of these weights requires a separate download of the base model.
|
57 |
|
58 |
```python
|
59 |
# Copyright 2024 Intel Corporation
|
@@ -135,4 +155,36 @@ inputs = processor(text=prompt, images=image, return_tensors="pt").to(device)
|
|
135 |
generate_ids = model.generate(**inputs, max_length=30)
|
136 |
output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
137 |
print(output)
|
138 |
-
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2 |
license: other
|
3 |
license_name: intel-research-use-license
|
4 |
license_link: LICENSE
|
5 |
+
tags:
|
6 |
+
- intel
|
7 |
+
- gaudi
|
8 |
+
- LLM
|
9 |
+
results:
|
10 |
+
- task:
|
11 |
+
type: Large Language Model
|
12 |
+
name: Large Language Model
|
13 |
+
metrics:
|
14 |
+
- type: GQA
|
15 |
+
name: GQA
|
16 |
+
value: 60.6138
|
17 |
+
- type: MMVP
|
18 |
+
name: MMVP
|
19 |
+
value: 36
|
20 |
+
- type: Pope Acc
|
21 |
+
name: Pope Acc
|
22 |
+
value: 87.33
|
23 |
+
- type: Pope F1
|
24 |
+
name: Pope F1
|
25 |
+
value: 86.5
|
26 |
+
- type: MMVet
|
27 |
+
name: MMVet
|
28 |
+
value: 31.9725
|
29 |
+
- type: ScienceQA
|
30 |
+
name: ScienceQA
|
31 |
+
value: 72.9797
|
32 |
+
- type: llavaw (1)
|
33 |
+
name: llavaw
|
34 |
+
value: 56.9
|
35 |
+
- type: llavaw (2)
|
36 |
+
name: llavaw
|
37 |
+
value: 61.9
|
38 |
+
- type: llavaw (3)
|
39 |
+
name: llavaw
|
40 |
+
value: 73.6
|
41 |
+
- type: llavaw (4)
|
42 |
+
name: llavaw
|
43 |
+
value: 65.7
|
44 |
+
|
45 |
+
library_name: transformers
|
46 |
+
pipeline_tag: image-text-to-text
|
47 |
---
|
48 |
|
49 |
+
## Model Details: LLaVA-llama-3-8B
|
50 |
|
51 |
+
`llava-llama-3-8b` is a large multimodal model (LMM) trained using the [LLaVA-v1.5 framework](https://arxiv.org/abs/2310.03744) with the 8-billion parameter [`meta-llama/Meta-Llama-3-8B-Instruct`](https://huggingface.co/meta-llama/Meta-Llama-3-8B) model as language backbone and the CLIP-based vision encoder.
|
52 |
|
53 |
+
| Model Details | Description |
|
54 |
+
| ----------- | ----------- |
|
55 |
+
| Authors | Intel: [Musashi Hinck*](https://huggingface.co/musashihinck), [Matthew L. Olson*](https://huggingface.co/matthewlyleolson), [Vasudev Lal](https://huggingface.co/vasudevlal) |
|
56 |
+
| Date | May 2024 |
|
57 |
+
| Version | 1 |
|
58 |
+
| Type | Large multimodal model (LMM) |
|
59 |
+
| Paper or Other Resources | [Improved Baselines with Visual Instruction Tuning](https://arxiv.org/abs/2310.03744) |
|
60 |
+
| License | [Intel Research Use License](https://huggingface.co/Intel/llava-llama-3-8b/blob/main/LICENSE) | All usage code is licensed Apache 2.0
|
61 |
+
| Questions or Comments | [Community Tab](https://huggingface.co/Intel/llava-llama-3-8b/discussions) and [Intel DevHub Discord](https://discord.gg/rv2Gp55UJQ)|
|
62 |
|
63 |
+
This model card was created by [Eduardo Alvarez](https://huggingface.co/eduardo-alvarez) and the authors listed above.
|
64 |
|
65 |
+
## Intended Use
|
66 |
|
67 |
+
| Intended Use | Description |
|
68 |
+
| ----------- | ----------- |
|
69 |
+
| Primary intended uses | The model has been finetuned for multimodal benchmark evaluations, but can also be used as a multimodal chatbot. |
|
70 |
+
| Primary intended users | Anyone using or evaluating multimodal models. |
|
71 |
+
| Out-of-scope uses | This model is not intended for uses that require high levels of factuality, high stakes situations, mental health or medical applications, generating misinformation or disinformation, impersonating others, facilitating or inciting harassment or violence, any use that could lead to the violation of a human right under the UN Declaration of Human Rights. |
|
72 |
|
|
|
73 |
|
74 |
+
### How to use
|
75 |
|
76 |
+
Please note, we only provide the trained weights difference and do not provide a copy of the base meta-llama/Meta-Llama-3-8B-Instruct model. Any use of these weights requires a separate download of the base model.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
77 |
|
78 |
```python
|
79 |
# Copyright 2024 Intel Corporation
|
|
|
155 |
generate_ids = model.generate(**inputs, max_length=30)
|
156 |
output = processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
157 |
print(output)
|
158 |
+
```
|
159 |
+
|
160 |
+
## Factors
|
161 |
+
|
162 |
+
| Factors | Description |
|
163 |
+
| ----------- | ----------- |
|
164 |
+
| Environment | Trained on a 4 node cluster with a total of 32 Gaudi 2 accelerators |
|
165 |
+
| Card Prompts | Model training and deployment on alternate hardware and software will change model performance |
|
166 |
+
|
167 |
+
## Training Data
|
168 |
+
|
169 |
+
The model was trained using the LLaVA-v1.5 data mixture. This is listed as follows:
|
170 |
+
|
171 |
+
- 558K filtered image-text pairs from LAION/CC/SBU, captioned by BLIP.
|
172 |
+
- 158K GPT-generated multimodal instruction-following data.
|
173 |
+
- 450K academic-task-oriented VQA data mixture.
|
174 |
+
- 40K ShareGPT data.
|
175 |
+
|
176 |
+
## Ethical Considerations
|
177 |
+
|
178 |
+
Intel is committed to respecting human rights and avoiding causing or contributing to adverse impacts on human rights. See [Intel’s Global Human Rights Principles](https://www.intel.com/content/dam/www/central-libraries/us/en/documents/policy-human-rights.pdf). Intel’s products and software are intended only to be used in applications that do not cause or contribute to adverse impacts on human rights.
|
179 |
+
|
180 |
+
| Ethical Considerations | Description |
|
181 |
+
| ----------- | ----------- |
|
182 |
+
| Data | The model was trained using the LLaVA-v1.5 data mixture as described above. |
|
183 |
+
| Human life | The model is not intended to inform decisions central to human life or flourishing. |
|
184 |
+
| Mitigations | No additional risk mitigation strategies were considered during model development. |
|
185 |
+
| Risks and harms | This model has not been assessed for harm or biases, and should not be used for sensitive applications where it may cause harm. |
|
186 |
+
| Use cases | - |
|
187 |
+
|
188 |
+
## Caveats and Recommendations
|
189 |
+
|
190 |
+
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. This model has not been assessed for harm or biases, and should not be used for sensitive applications where it may cause harm.
|