AI & ML interests

Machine learning, deep learning, generative AI, LLMs

Recent Activity

salma-remyx 
updated a Space 1 day ago
salma-remyx 
in remyxai/SpaceOm 24 days ago

How to Train the Model?

1
#1 opened 24 days ago by
qhz991029
salma-remyx 
posted an update 24 days ago
salma-remyx 
posted an update about 1 month ago
view post
Post
1584
When multiple benchmarks yield conflicting model rankings, how do you know which model to trust?

In this substack, we explore that question in the context of spatial reasoning capabilities as seen from the perspective of 3 new benchmarks.

Read more: https://remyxai.substack.com/p/benchmark-fusion
salma-remyx 
posted an update 3 months ago
view post
Post
1868
SpaceThinker-Qwen2.5VL-3B shows a 3B VLM can compete with closed, frontier APIs in quantitative spatial reasoning, a key capability for embodied AI applications like drones and robotics.

Check out how it stacks up against Gemini and OpenAI on Q-Spatial-Bench in the ModelCard. Includes .gguf, colab quickstart, docker images.


SpaceThinker adopts the Qwen2.5VL-3B architecture, fine-tuned on the SpaceThinker dataset of synthetic spatial reasoning traces, created with VQASynth

This model builds upon the SpaceLLaVA series of VLMs finetuned for enhanced spatial reasoning using synthetic data by adding test-time compute for multimodal thinking.

Model: remyxai/SpaceThinker-Qwen2.5VL-3B
Dataset: remyxai/SpaceThinker
Space: remyxai/SpaceThinker-Qwen2.5VL-3B
Code: https://github.com/remyxai/VQASynth
Discussion: open-r1/README#10

  • 1 reply
·