SWAP_v2_MATH500_Disc_Llama3-8B-LoRA

SWAP_v2 MATH500 Discriminator (Llama3-8B-LoRA)

This repository contains the MATH500 discriminator (Llama3-8B-LoRA) trained with SWAP_v2 .

Model details

• Role: Discriminator
• Dataset: MATH500
• Base model: meta-llama/Meta-Llama-3-8B-Instruct
• Adapter type: LoRA
• LoRA rank (r): 16
• LoRA alpha: 32
• LoRA dropout: 0.1
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Bias: "none"

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model_id = "meta-llama/Meta-Llama-3-8B-Instruct"
adapter_id = "sxiong/SWAP_v2_MATH500_Disc_Llama3-8B-LoRA"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)

For additional information and implementation details, please refer to the SWAP GitHub repository.

Citation

@inproceedings{xiong2025deliberate,
  title={Deliberate reasoning in language models as structure-aware planning with an accurate world model},
  author={Xiong, Siheng and Payani, Ali and Yang, Yuan and Fekri, Faramarz},
  booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  pages={31900--31931},
  year={2025}
}