VIM 27B trained on VI AMO v2

This LoRA adapter jointly predicts ongoing action and ongoing mission from visual teleoperation history. It is the final checkpoint from the completed VI AMO v2 training run used for the VIM (Ours) evaluation.

  • Base model: unsloth/Qwen3.5-27B, revision 68443620da0c3847c1ba5042b5dcf1e78c42bfba.
  • Dataset: VI AMO v2, with 6,600 episodes, 13,653 actions, and 9,702 missions.
  • Training: one epoch, 4,306 optimizer steps, LoRA rank and alpha 16.
  • Evaluation: 1,091 test actions at nine observation cutoffs, yielding 9,819 predictions.
  • Weight SHA-256: 4759b2b337c8003114e8d955776118fd5d9dc9ebc82a36236690335c8f0d8e0f.

See checkpoint.json for provenance and SHA256SUMS for file checksums. The adapter and processor files are copied unchanged from the training output. The base model must be downloaded separately.

Download this adapter from the VIM repository root:

hf download viamodelanddataset/vim27b_lora_v2 \
  --local-dir checkpoints/vim27b_lora_v2
(cd checkpoints/vim27b_lora_v2 && sha256sum --check SHA256SUMS)

Follow the installation, base-model, and inference instructions and use --model checkpoints/vim27b_lora_v2.

The adapter weights, configuration, processor, and provenance are identical to the v2 checkpoint previously included in the VIM code repository. Only the model card and its checksum were updated for hosting on Hugging Face.

Downloads last month
22
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for vi-amo/vim27b_lora_v2

Base model

Qwen/Qwen3.5-27B
Adapter
(27)
this model

Dataset used to train vi-amo/vim27b_lora_v2