cronusvla_modified_ckpt_post-train

Model Card for CronusVLA

Weights

step2/checkpoints/reshape_embedding_step2.pt:

• (1) This checkpoint serves as the starting point for fine-tuning on the Libero benchmark.
• (2) It is a post-trained model derived from CronusVLA-7B, configured with a two-step frame (history length = 1).
• (3) During the original post-training phase, the model was trained to predict future actions with a chunk size of 16. To ensure compatibility with the Libero setting of OpenVLA-oft (chunk = 8), we modified the position embeddings responsible for action chunking in the original checkpoint.
• (4) Note: This checkpoint is not directly evaluable and is intended only as a fine-tuning initialization.

step4/checkpoints/reshape_embedding_step4.pt:

• (1) This checkpoint serves as the starting point for fine-tuning on the LIBERO benchmark.
• (2) It is a post-trained model derived from CronusVLA-7B, configured with a four-step frame (history length = 3).
• (3) During the original post-training phase, the model was trained to predict future actions with a chunk size of 16. To ensure compatibility with the Libero setting of OpenVLA-oft (chunk = 8), we modified the position embeddings responsible for action chunking in the original checkpoint.
• (4) Note: This checkpoint is not directly evaluable and is intended only as a fine-tuning initialization.

If you want to use these checkpoint to finetune, please follow the instruction of CronusVLA.