llama3_8b_truth_rm_full

llama3_8b_truth_rm_full

This model is a Truthfulness Process Reward Model (PRM) introduced in the paper Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards.

It is part of the MAHALO (Multi-Action-Head ALignment with PRM-guided DecOding) framework, which standardizes alignment across multiple objectives including math reasoning and human preferences. This specific checkpoint provides step-level or sequence-level supervision for the truthfulness dimension within the Human Values domain.

• Paper: Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards
• Repository: https://github.com/pearls-lab/multiobj-align

Model Description

The model is a fine-tuned version of Jennny/llama3_8b_sft_ultrafb based on the Llama-3-8B architecture. It is trained to serve as a reward model to evaluate and encourage truthful responses in open-domain conversations, using the UltraFeedback dataset.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 16
• total_train_batch_size: 512
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 1

Training results

Framework versions

• Transformers 4.45.2
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3

Citation

If you find this work useful, please consider citing:

@article{shen2025simultaneous,
  title={Simultaneous Multi-objective Alignment Across Verifiable and Non-verifiable Rewards},
  author={Shen, Yiran and Xia, Yu and Chang, Jonathan and Ammanabrolu, Prithviraj},
  journal={arXiv preprint arXiv:2510.01167},
  year={2025},
  url={https://arxiv.org/abs/2510.01167}
}