Qwen2.5-VL-3B-Instruct-Action

Qwen2.5-VL-3B-Instruct with action tokens

This repository is a Qwen2.5-VL-3B-Instruct initializer whose tokenizer and text vocabulary have been extended with 2,048 tokens named <robot_action_0> through <robot_action_2047>. It is intended as a starting point for StarVLA frameworks that discretize robot actions as tokens.

The packaged config.json identifies Qwen2_5_VLForConditionalGeneration, a 2,048-dimensional, 36-layer text backbone originating from Qwen2.5-VL-3B-Instruct.

Model and input contract

The extension was produced with starVLA/model/modules/vlm/tools/add_qwen_special_tokens.

Training and evaluation

No additional VLM or robot-policy fine-tuning is reported for this repository. The pre-existing Qwen2.5-VL parameters are retained apart from resizing the token embedding/output vocabulary for the new entries. The added action-token embeddings therefore should be treated as initialization, not as learned action semantics.

No robot benchmark result is associated with this repository.

Loading

Use the model and tokenizer from the same repository so their vocabulary sizes and token IDs remain aligned:

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

repo_id = "StarVLA/Qwen2.5-VL-3B-Instruct-Action"
processor = AutoProcessor.from_pretrained(repo_id)
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    repo_id,
    torch_dtype="auto",
    device_map="auto",
)

Downstream training must define the action quantization/codebook that maps continuous robot actions to the 2,048 token indices.

Evidence and limitations

• config.json verifies the Qwen2.5-VL architecture and 3B-class hidden dimensions; added_tokens.json verifies all 2,048 action-token entries.
• This repository does not define normalization, action dimension, action horizon, robot embodiment, dataset, controller, or evaluation protocol.
• Token IDs are repository-specific. Mixing this tokenizer with the unextended base model or another action-token checkpoint can silently corrupt decoding.
• The model inherits the base model license, acceptable-use requirements, biases, and failure modes.