---
library_name: transformers
tags: []
pipeline_tag: image-text-to-text
license: apache-2.0
datasets:
- LangAGI-Lab/WebPRMCollection_preference_pair
language:
- en
---

# Model Card for Web-Shepherd

Web-Shepherd is the first process reward model (PRM) designed specifically for web agents, as presented in the paper [Web-Shepherd: Advancing PRMs for Reinforcing Web Agents](https://arxiv.org/abs/2505.15277). It evaluates trajectories at the step level to provide interpretable and cost-efficient feedback for both learning and inference-time decision making in web navigation tasks.

## Model Details

*   **Developed by:** [More Information Needed]
*   **Model type:** Language Model
*   **License:** apache-2.0
*   **Finetuned from model:** Qwen3

### Model Sources

- **Repository:** https://github.com/LangAGI-Lab/WebShepherd
- **Paper:** https://arxiv.org/abs/2505.15277
- **Dataset:** https://huggingface.co/datasets/LangAGI-Lab/WebPRMCollection_preference_pair

## Uses

### Direct Use

This model can be used to assess web navigation trajectories in a step-level.

### Downstream Use

The model can be fine-tuned for web navigation tasks.

## Training Details

### Training Data

The model has been trained on the WebPRM Collection, a large-scale dataset with 40K step-level preference pairs and annotated checklists spanning diverse domains and difficulty levels.

## Evaluation

The model was evaluated on the WebRewardBench, the first meta-evaluation benchmark for evaluating PRMs.