nvidia/Llama-3.3-Nemotron-70B-Feedback

Model Overview

Description:

Llama-3.3-Nemotron-70B-Feedback is a large language model that leverages Meta-Llama-3.3-70B-Instruct as the foundation and is fine-tuned using Supervised Finetuning to provide feedback regarding the helpfulness of LLM generated responses to user queries.

This model is ready for commercial use.

License/Terms of Use:

GOVERNING TERMS: Use of this model is governed by the NVIDIA Open Model License . Additional Information: Llama 3.3 Community License Agreement. Built with Llama.

Arena Hard LeaderBoard

As of 18 Mar 2025, augmenting models with the Feedback-Edit Inference Time Scaling (ITS) approach leads to the highest performance on Arena Hard.

The Feedback-Edit Inference Time Scaling system comprise of the following models:

1. Llama-3.3-Nemotron-70B-Feedback
2. Llama-3.3-Nemotron-70B-Edit
3. Llama-3.3-Nemotron-70B-Select

Model	Arena Hard (95% CI)
Llama-3.3-Nemotron-Super-49B-v1 + Feedback-Edit ITS	93.4 (-1.1, 1.0)
Llama-3.1-Nemotron-70B-Instruct + Feedback-Edit ITS	92.7 (-1.2, 0.9)
o1-mini-2024-09-12	92.0 (-1.2, 1.0)
o1-preview-2024-09-12	90.4 (-1.1, 1.3)
Llama-3.3-Nemotron-Super-49B-v1	88.3 (-1.6, 1.6)
claude-3-5-sonnet-20241022	85.2 (-1.4, 1.6)
Llama-3.1-Nemotron-70B-Instruct	84.9 (-1.7, 1.8)

Use Case:

Llama-3.3-Nemotron-70B-Feedback provides feedback regarding the helpfulness of LLM generated responses to user queries, for users who are interested in improving performance through Inference-Time-Scaling for general-domain, open-ended tasks.

Release Date:

03/18/2025

References(s):

• Dedicated Feedback and Edit Models Empower Inference-Time Scaling for Open-Ended General-Domain Tasks
• HelpSteer2-Preference
• SteerLM method
• HelpSteer
• HelpSteer2
• The future of AI: Built with Llama
• Meta's Llama 3.3 Webpage
• Meta's Llama 3.3 Model Card

Model Architecture:

Architecture Type: Transformer
Network Architecture: Llama 3.3

We developed this model using Llama-3.3-70B-Instruct as its foundation. This model contains 70 billion parameters.

Input:

Input Type(s): Text
Input Format: String
Input Parameters: One Dimensional (1D)
Other Properties Related to Input: Max of 128k tokens

Output:

Output Type(s): Text
Output Format: String
Output Parameters: One Dimensional (1D)
Other Properties Related to Output: Max of 4k tokens

Software Integration:

Runtime Engine(s):

• [NeMo - 24.05.llama.3.1]
  

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
  
• NVIDIA Hopper
  
• NVIDIA Turing
  Supported Operating System(s): Linux
  

Quick Start

You can use the model using HuggingFace Transformers library with 2 or more 80GB GPUs (NVIDIA Ampere or newer) with at least 150GB of free disk space to accomodate the download.

This code has been tested on Transformers v4.45.0, torch v2.3.0a0+40ec155e58.nv24.3 and 2 A100 80GB GPUs, but any setup that supports meta-llama/Llama-3.1-70B-Instruct should support this model as well. If you run into problems, you can consider doing pip install -U transformers.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "nvidia/Llama-3.3-Nemotron-70B-Feedback"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)


def generate_feedback(messages, model, tokenizer, temperature=0.7):
    tokenized_message = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True)
    response_token_ids = model.generate(tokenized_message['input_ids'].cuda(),attention_mask=tokenized_message['attention_mask'].cuda(),  max_new_tokens=128, pad_token_id = tokenizer.eos_token_id, num_return_sequences=1, temperature=temperature)
    generated_tokens =response_token_ids[:, len(tokenized_message['input_ids'][0]):]
    generated_text = tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
    return generated_text

prompt = "What is the distance between the Earth and the Sun?"
response = "The distance from Earth to the Sun is 93 million miles"

messages = [
        {"role": "user", "content": prompt},
        {"role": "assistant", "content": response},
        {"role": "user", "content": "Evaluate the response to the previous prompt in terms of how helpful it is overall. Start the evaluation with the statement - The response is {not / slightly / partially / mostly / perfectly} helpful. Then provide a brief explanation of the evaluation in 2 to 10 sentences."}
    ]

feedback = generate_feedback(messages, model, tokenizer, temperature=0.7)
print(feedback)

## Illustrative example - feedback generated might not be identical since temperature sampling is used

# The response is partially helpful. It provides a concise answer to the prompt. However, the lack of additional information or context limits its usefulness. It could have been more informative by including the average distance in astronomical units (AU) and explaining the variation in distance due to the elliptical orbit.

Model Version:

v1.0

Training and Testing Datasets:

Training Datasets:

Dataset Name: HelpSteer3 Dataset Link: https://huggingface.co/datasets/nvidia/HelpSteer3

Data Collection Method by dataset

• [Hybrid: Human, Synthetic]
  

Labeling Method by dataset

• [Human]
  

Properties:

• 77,564 prompt-responses, each annotated with up to 3 annotations of free-text feedback (each being 50-250 words long) elaborating upon the overall helpfulness of the response.

Testing Datasets:

Dataset Name: HelpSteer3 Dataset Link: https://huggingface.co/datasets/nvidia/HelpSteer3

Data Collection Method by dataset

• [Hybrid: Human, Synthetic]
  

Labeling Method by dataset

• [Human]
  

Properties:

• 4,078 prompt-responses, each annotated with up to 3 annotations of free-text feedback (each being 50-250 words long) elaborating upon the overall helpfulness of the response.

Inference:

Engine: Triton
Test Hardware: H100, A100 80GB, A100 40GB

Limitations:

The model was trained on data that contains toxic language, unsafe content, and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their supporting model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse. For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards.
Please report security vulnerabilities or NVIDIA AI Concerns here.

Citation

If you find this model useful, please cite the following work:

@misc{wang2025dedicatedfeedbackeditmodels,
      title={Dedicated Feedback and Edit Models Empower Inference-Time Scaling for Open-Ended General-Domain Tasks}, 
      author={Zhilin Wang and Jiaqi Zeng and Olivier Delalleau and Daniel Egert and Ellie Evans and Hoo-Chang Shin and Felipe Soares and Yi Dong and Oleksii Kuchaiev},
      year={2025},
      eprint={2503.04378},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.04378},
}