Qwen2.5-CascadeRL-RM-72B

Description

Qwen2.5-CascadeRL-RM-72B is a reward model that is initialized with Qwen2.5-72B-Instruct and is fine-tuned using the Bradley-Terry objective to predict the human preference of LLM generation. It is used in Reinforcement Learning from Human Feedback (RLHF) stage in Nemotron-Cascade model family: Nemotron-Cascade-8B, Nemotron-Cascade-8B-Thinking, and Nemotron-Cascade-14B-Thinking.

Given a conversation between a human and an assistant, the reward model will give a human preference score for the final assistant turn.

For the training details, please refer to the technical report.

Usage Recommendations

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "nvidia/Qwen2.5-CascadeRL-RM-72B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",  
    low_cpu_mem_usage=True,
    torch_dtype=torch.bfloat16,
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
model.eval()

prompt = "Hello! How are you?"
response = "I am fine! Thanks for asking. How are you?"
messages = [{"role":"user","content":prompt}, {"role":"assistant","content":response}]
batch = tokenizer.apply_chat_template(
    messages, tokenize=True, add_generation_prompt=False,
    return_tensors="pt", return_dict=True
)
with torch.inference_mode():
    out = model(**batch, use_cache=False)
print(out.logits[0, -1, 0].item())

RewardBench

Model	Overall	Chat	Chat Hard	Safety	Reasoning
Qwen2.5-CascadeRL-RM-72B	95.15	98.60	89.69	93.92	98.40

Release Date

Dec 31, 2025

License

Your use of this model is governed by the NVIDIA Open Model License.

Citation

@article{Nemotron_Cascade_Scaling_Cascaded_Reinforcement_Learning,
  title={Nemotron-Cascade: Scaling Cascaded Reinforcement Learning for General-Purpose Reasoning Models},
  author={Wang, Boxin and Lee, Chankyu and Lee, Nayeon and Lin, Sheng-Chieh and Dai, Wenliang and Chen, Yang and Chen, Yangyi and Yang, Zhuolin and Liu, Zihan and Shoeybi, Mohammad and Catanzaro, Bryan and Ping, Wei},
  year={2025}
}