KITE-7B-Instruct

KITE-7B-Instruct is a fine-tuned version of Qwen2.5-VL-7B-Instruct for VLM-based robot failure analysis, released as part of the KITE paper (ICRA 2026).

This checkpoint contains the full merged weights (base + LoRA adapter), ready for direct inference with no additional merge step.

Model Details

Base model	Qwen/Qwen2.5-VL-7B-Instruct
Parameters	~7B
Fine-tuning	QLoRA (4-bit NF4) on RoboFAC textual + multimodal tasks
Architecture	Qwen2.5-VL (vision-language, conditional generation)
License	Apache 2.0 (same as base model)

Usage

from transformers import AutoProcessor, AutoModelForVision2Seq

model_id = "m80hz/KITE-7B-Instruct"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForVision2Seq.from_pretrained(model_id, device_map="auto", trust_remote_code=True)

Or serve it with vLLM for OpenAI-compatible inference:

python -m vllm.entrypoints.openai.api_server --model m80hz/KITE-7B-Instruct

Then use the KITE pipeline to run failure analysis:

python -m kite.cli \
  --model_name m80hz/KITE-7B-Instruct \
  --model_url http://127.0.0.1:8000/v1 \
  --dataset_folder ./datasets/robofac/simulation_data \
  --test_file ./datasets/robofac/test_qa_sim/test_detect_identify_locate.json \
  --out_dir ./outputs/kite_run

Usage

@inproceedings{hosseinzadeh2025kite,
  title     = {KITE: Keyframe-Indexed Tokenized Evidence for VLM-Based Robot Failure Analysis},
  author    = {Hosseinzadeh, Mehdi and Wong, King Hang and Dayoub, Feras},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2026}
}