verification-1.5b / README.md
nielsr's picture
nielsr HF Staff
Add model card for Budget-Aware Test-Time Scaling via Discriminative Verification
04fa90e verified
|
raw
history blame
2.83 kB
metadata
pipeline_tag: text-ranking
library_name: transformers
license: apache-2.0

Budget-Aware Test-Time Scaling via Discriminative Verification

This repository contains the verification-1.5b model, implementing the discriminative verifier described in the paper: Budget-Aware Test-Time Scaling via Discriminative Verification

📃 [Paper on arXiv] • 💻 [GitHub] • 🤗 [Hugging Face Collection]

Model Description

This model is a discriminative verifier designed for budget-aware test-time scaling in large language models on complex reasoning tasks. While state-of-the-art approaches often employ generative verifiers, this model shifts focus to a more budget-aware paradigm: discriminative verification. Empirical analysis demonstrates that combining discriminative verifiers with self-consistency creates a powerful and efficient test-time scaling mechanism, surpassing state-of-the-art generative verification under a fixed compute budget. This approach is not only a "free" upgrade over self-consistency but also a more effective and efficient alternative to costly generative techniques for practical, real-world applications.

Usage

You can use this model to score candidate solutions for various evaluation datasets.

First, ensure you have the necessary environment set up by installing the package:

conda create -n verification python=3.10
conda activate verification
pip install -e . # This installs the 'verification' package and its dependencies

Then, you can run verification on an evaluation dataset (e.g., aime2024, aime2025, livebench-math, or gpqa) with the trained verifier using the run_judge_hf.py script:

python scripts/run_judge_hf.py \
    --model_name "WangResearchLab/verification-1.5b" \
    --dataset_name "verification-evaluation-data" \
    --dataset_split "validation" \
    --save_path "evals/verification-1.5b/validation-eval.jsonl" \
    --num_gpus 8

This command will score candidate solutions with the WangResearchLab/verification-1.5b model and save the evaluation results to the specified path.

Citation

If you find this work useful, please consider citing the paper:

@article{montgomery2025budget,
  title={Budget-Aware Test-Time Scaling via Discriminative Verification},
  author={Montgomery, Kyle and Tan, Sijun and Chen, Yuqi and Zhuang, Siyuan and Zhang, Tianjun and Popa, Raluca Ada and Wang, Chenguang},
  journal={arXiv preprint arXiv:2510.14913},
  year={2025}
}