Phi-4 Technical Report
Paper
โข
2412.08905
โข
Published
โข
101
Fatih C. Akyon
fcakyon
AI & ML interests
multi-modal learning, video understanding
Recent Activity
upvoted
a
paper
3 days ago
Phi-4 Technical Report
new activity
7 days ago
thwri/CogFlorence-2:๐ฉ Report: Not working
liked
a Space
10 days ago
thwri/CogFlorence-2
Organizations
fcakyon's activity
upvoted
a
paper
3 days ago
๐ฉ Report: Not working
2
#1 opened 10 days ago
by
fcakyon
liked
a
Space
10 days ago
add_confidence_score
2
#56 opened 6 months ago
by
haipingwu
broken link in the model card
1
#99 opened 10 days ago
by
fcakyon
reacted to
merve's
post with ๐ค๐๐
15 days ago
Post
3296
Apollo is a new family of open-source video language models by Meta, where 3B model outperforms most 7B models and 7B outperforms most 30B models ๐งถ
โจ the models come in 1.5B https://huggingface.co/Apollo-LMMs/Apollo-1_5B-t32, 3B https://huggingface.co/Apollo-LMMs/Apollo-3B-t32 and 7B https://huggingface.co/Apollo-LMMs/Apollo-7B-t32 with A2.0 license, based on Qwen1.5 & Qwen2
โจ the authors also release a benchmark dataset https://huggingface.co/spaces/Apollo-LMMs/ApolloBench
The paper has a lot of experiments (they trained 84 models!) about what makes the video LMs work โฏ๏ธ
Try the demo for best setup here https://huggingface.co/spaces/Apollo-LMMs/Apollo-3B
they evaluate sampling strategies, scaling laws for models and datasets, video representation and more!
> The authors find out that whatever design decision was applied to small models also scale properly when the model and dataset are scaled ๐ scaling dataset has diminishing returns for smaller models
> They evaluate frame sampling strategies, and find that FPS sampling is better than uniform sampling, and they find 8-32 tokens per frame optimal
> They also compare image encoders, they try a variation of models from shape optimized SigLIP to DINOv2
they find google/siglip-so400m-patch14-384 to be most powerful ๐ฅ
> they also compare freezing different parts of models, training all stages with some frozen parts give the best yield
They eventually release three models, where Apollo-3B outperforms most 7B models and Apollo 7B outperforms 30B models ๐ฅ
โจ the models come in 1.5B https://huggingface.co/Apollo-LMMs/Apollo-1_5B-t32, 3B https://huggingface.co/Apollo-LMMs/Apollo-3B-t32 and 7B https://huggingface.co/Apollo-LMMs/Apollo-7B-t32 with A2.0 license, based on Qwen1.5 & Qwen2
โจ the authors also release a benchmark dataset https://huggingface.co/spaces/Apollo-LMMs/ApolloBench
The paper has a lot of experiments (they trained 84 models!) about what makes the video LMs work โฏ๏ธ
Try the demo for best setup here https://huggingface.co/spaces/Apollo-LMMs/Apollo-3B
they evaluate sampling strategies, scaling laws for models and datasets, video representation and more!
> The authors find out that whatever design decision was applied to small models also scale properly when the model and dataset are scaled ๐ scaling dataset has diminishing returns for smaller models
> They evaluate frame sampling strategies, and find that FPS sampling is better than uniform sampling, and they find 8-32 tokens per frame optimal
> They also compare image encoders, they try a variation of models from shape optimized SigLIP to DINOv2
they find google/siglip-so400m-patch14-384 to be most powerful ๐ฅ
> they also compare freezing different parts of models, training all stages with some frozen parts give the best yield
They eventually release three models, where Apollo-3B outperforms most 7B models and Apollo 7B outperforms 30B models ๐ฅ
-
- 6 replies
Great share @merve ๐ฏApollo links in your post not working, giving 404 for me ๐ค
liked
3
datasets
about 2 months ago
Inherit from GenerationMixin
#22 opened 2 months ago
by
fcakyon
updated
a
model
2 months ago
Inherit from GenerationMixin
3
#80 opened 2 months ago
by
Link161
updated
4
Spaces
3 months ago
liked
a
Space
3 months ago