AIcell/Guava-4B
AIcell/Guava-4B is a 4.5 billion parameter vision-language model based on the Qwen3.5-VL architecture, developed by AIcell. It is specifically fine-tuned on the Guava SFT dataset, which comprises 2,268 episodes across 16 tabletop robot manipulation tasks. This model excels at understanding and generating responses for robotic manipulation scenarios, particularly those involving object interaction and spatial reasoning on tabletops, with a context length of 32,768 tokens.
Loading preview...
Overview
AIcell/Guava-4B is a specialized 4.5 billion parameter vision-language model built upon the Qwen3_5ForConditionalGeneration architecture. It is uniquely fine-tuned on the Guava SFT data, a comprehensive dataset featuring 2,268 episodes across 16 distinct tabletop robot manipulation tasks. This model is designed for applications requiring visual understanding and textual interaction in robotic control environments.
Key Capabilities
- Vision-Language Integration: Combines a
Qwen3_5ForConditionalGenerationarchitecture with a 24-layer ViT for robust image and text processing. - Robot Manipulation Focus: Specifically trained on the AIcell/Guava-4B-SFT-data dataset, making it highly proficient in understanding and executing commands related to robot manipulation.
- High Context Length: Supports a maximum context length of 262,144 positions (32,768 tokens), allowing for complex sequences of instructions and observations.
- Precision: Utilizes
bfloat16precision for efficient computation.
Good for
- Robotics Research: Ideal for researchers working on robot learning, imitation learning, and visual-language navigation in tabletop environments.
- Task Planning: Suitable for developing agents that can interpret visual scenes and generate action plans for tasks such as
cube_stack,pick_up_orange, andpush_basket. - Simulated Environments: Can be used to train and evaluate robotic policies in simulated tabletop scenarios, leveraging its specialized training data.