🔍 xVerify-3B-Ia

xVerify is an evaluation tool fine-tuned from a pre-trained large language model, designed specifically for objective questions with a single correct answer. It accurately extracts the final answer from lengthy reasoning processes and efficiently identifies equivalence across different forms of expressions.

This model was introduced in the paper xVerify: Efficient Answer Verifier for Reasoning Model Evaluations.

---

✨ Key Features

📊 Broad Applicability

Suitable for various objective question evaluation scenarios including math problems, multiple-choice questions, classification tasks, and short-answer questions.

⛓️ Handles Long Reasoning Chains

Effectively processes answers with extensive reasoning steps to extract the final answer, regardless of complexity.

🌐 Multilingual Support

Primarily handles Chinese and English responses while remaining compatible with other languages.

🔄 Powerful Equivalence Judgment

• ✓ Recognizes basic transformations like letter case changes and Greek letter conversions
• ✓ Identifies equivalent mathematical expressions across formats (LaTeX, fractions, scientific notation)
• ✓ Determines semantic equivalence in natural language answers
• ✓ Matches multiple-choice responses by content rather than just option identifiers

---

🚀 Usage

For detailed instructions on installation and batch evaluation using the xVerify framework, please refer to the official GitHub repository.

Since this is a Llama-based model, you can also use it directly with the transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "IAAR-Shanghai/xVerify-3B-Ia"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

# Your input prompt logic here

---

📚 Citation

@article{xVerify,
      title={xVerify: Efficient Answer Verifier for Reasoning Model Evaluations}, 
      author={Ding Chen and Qingchen Yu and Pengyuan Wang and Wentao Zhang and Bo Tang and Feiyu Xiong and Xinchi Li and Minchuan Yang and Zhiyu Li},
      journal={arXiv preprint arXiv:2504.10481},
      year={2025},
}