VLMPed-wo-CoT

A LoRA fine-tuned version of Qwen2.5-VL-3B-Instruct for pedestrian crossing intention prediction, trained without Chain-of-Thought supervision.

This model is part of the ECE 228 final project at UCSD (Spring 2026): "How Do Vision Language Models Utilize Multi-Frame Temporal Information for Pedestrian Intention Prediction?"

Project repository: ece228_VLMPed-CoT

---

Model Details

• Developed by: chiawen0104
• Model type: Vision-Language Model (LoRA fine-tuned)
• Finetuned from: Qwen/Qwen2.5-VL-3B-Instruct
• Task: Pedestrian crossing intention prediction (binary: cross / not cross)
• Training datasets: JAAD, PIE
• Framework: PEFT 0.15.1

---

How to Get Started

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
from peft import PeftModel

base_model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    "Qwen/Qwen2.5-VL-3B-Instruct"
)
model = PeftModel.from_pretrained(base_model, "chiawen0104/VLMPed-wo-CoT")
processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-3B-Instruct")

---

Training Details

• Base model: Qwen2.5-VL-3B-Instruct
• Fine-tuning method: LoRA (via PEFT)
• Training regime: bf16 mixed precision
• Training data: JAAD and PIE pedestrian crossing intention datasets
• CoT supervision: None (direct prediction without chain-of-thought)

---

Intended Use

This model takes multi-frame pedestrian images as input and predicts whether a pedestrian intends to cross the street. It is intended for research purposes in autonomous driving and pedestrian behavior analysis.

---

Differences from VLMPed-CoT

	VLMPed-CoT	VLMPed-wo-CoT
CoT supervision	✅	❌
Direct prediction	✅	✅

---

Reference

• Original Paper: VLMPed-CoT: A large vision-language model with a chain-of-thought mechanism for pedestrian crossing intention prediction
• Original implementation: lyc2121/VLMPed-CoT-for-Pedestrian-Crossing-Intention-Prediction
• Companion model: chiawen0104/VLMPed-CoT

Framework versions

• PEFT 0.15.1