CounterVid-Qwen2.5-VL-3B-LoRA

CounterVid — Qwen2.5-VL-3B

This repository contains the 3B CounterVid-aligned model from CounterVid: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models (EMNLP 2026).

CounterVid improves grounding in video-language models by combining textual and visual preferences generated from controlled counterfactual videos. The released weights are a standalone merge of the trained LoRA adapter, the fully trained visual merger, and the original Qwen2.5-VL base model.

Loading

import torch
from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration

model_id = "aimagelab/CounterVid-Qwen2.5-VL-3B-LoRA"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
).eval()
processor = AutoProcessor.from_pretrained(model_id)

For complete video inference and evaluation examples, see the CounterVid repository and project page.

Training summary

• Base model: Qwen/Qwen2.5-VL-3B-Instruct
• Training data: CounterVid (26,167 synthetic visual and textual preference pairs)
• Objective: MixDPO / PaMi-VDPO with CounterVid anchors
• LoRA was merged before release; PEFT is not required at inference time
• Vision encoder frozen during preference optimization

Intended use and limitations

This is a research model for video-language understanding, with a focus on action recognition and temporal reasoning. It inherits the capabilities, limitations, biases, and license conditions of its base model. It may still produce incorrect or unsupported answers and should not be used as the sole basis for high-stakes decisions.

Citation

@inproceedings{poppi2026countervid,
  title     = {{CounterVid}: Counterfactual Video Generation for Mitigating Action and Temporal Hallucinations in Video-Language Models},
  author    = {Poppi, Tobia and Uzkent, Burak and Garg, Amanmeet and Porto, Lucas and Kessler, Garin and Yang, Yezhou and Cornia, Marcella and Baraldi, Lorenzo and Cucchiara, Rita and Schiffers, Florian},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}

This work was conducted during Tobia Poppi's Applied Scientist internship with the Amazon Prime Video team in Seattle.