You can deploy this model and start earning money today!

denkiwakame/
Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3

Name: Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3
Availability: InStock

vision-language-action

knowledge-distillation

qwen2.5-vl

Qwen/Qwen2.5-VL-3B-Instruct

finetune

cc-by-nc-4.0

endpoints_compatible

Views

No views yet

Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3

VLA-0 checkpoint trained with token-level knowledge distillation.

Student: denkiwakame/Qwen3.5-0.8B-FFT-VLA0
Teacher: denkiwakame/Qwen3.5-4B-FFT-VLA0
Dataset: LIBERO (via RoboVerse)
Method: KL divergence on digit token logits at action positions

Quick Start

1. Download

bash

pip install huggingface_hub
huggingface-cli download denkiwakame/Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3 --local-dir ./Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3

2. Load with transformers

python

import pickle
import torch
from transformers import AutoModelForCausalLM, AutoProcessor

ckpt_dir = "./Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3/model_final"

model = AutoModelForCausalLM.from_pretrained(ckpt_dir, torch_dtype=torch.bfloat16, device_map="auto")
processor = AutoProcessor.from_pretrained(ckpt_dir)

with open("./Qwen3.5-0.8B-FFT-VLA0-KD-T4B-CE0.3/dataset_stats.pkl", "rb") as f:
    dataset_stats = pickle.load(f)

Files

File	Description
`model_final/model.safetensors`	Distilled model weights
`model_final/config.json`	Model configuration
`model_final/tokenizer.json`	Tokenizer
`dataset_stats.pkl`	Action normalization statistics (required for inference)
`config.yaml`	Training configuration

Training Details

Training Config

License

CC-BY-NC-4.0 (following the upstream VLA-0 license). Subject to Qwen License for the base model.