Hugging Face's logo

nvidia
/
Llama-3.1-Nemotron-70B-Reward-HF

Safetensors
English
llama
nvidia
llama3.1
reward model
Model card Files Community
4
Llama-3.1-Nemotron-70B-Reward-HF / README.md
zhilinw's picture
zhilinw
Update README.md
cfff9a6 verified
preview code
|
raw
history blame contribute delete
12.5 kB
metadata
license: llama3.1
language:
  - en
inference: false
fine-tuning: false
tags:
  - nvidia
  - llama3.1
  - reward model
datasets:
  - nvidia/HelpSteer2
base_model: meta-llama/Llama-3.1-70B-Instruct

Model Overview

Description:

Llama-3.1-Nemotron-70B-Reward is a large language model customized using developed by NVIDIA to predict the quality of LLM generated responses. Specifically, it has been trained using a Llama-3.1-70B-Instruct Base on a novel approach combining the strength of Bradley Terry and SteerLM Regression Reward Modelling.

Given a English conversation with multiple turns between user and assistant (of up to 4,096 tokens), it rates the quality of the final assistant turn using a reward score.

For the same prompt, a response with higher reward score has higher quality than another response with a lower reward score, but the same cannot be said when comparing the scores between responses to different prompts.

Llama-3.1-Nemotron-70B-Reward-HF has been converted from Llama-3.1-Nemotron-70B-Reward to support it in the HuggingFace Transformers codebase. Please note that evaluation results might be slightly different from the Llama-3.1-Nemotron-70B-Reward as evaluated in NeMo-Aligner, which the evaluation results below are based on.

Try hosted inference for free at build.nvidia.com - it comes with an OpenAI-compatible API interface and simply signing up gets you 100k free API calls to this model.

Using this reward model for RLHF (specifically, REINFORCE), we were able to tune a Llama-3.1-70B-Instruct model to reach AlpacaEval 2 LC of 57.6, Arena Hard of 85.0 and GPT-4-Turbo MT-Bench of 8.98, which are known to be predictive of LMSys Chatbot Arena Elo This Instruct model is available at Llama-3.1-Nemotron-70B-Instruct as .nemo model and Llama-3.1-Nemotron-70B-Instruct-HF as a HF Transformers model.

As of 1 Oct 2024, this model is #1 on all three automatic alignment benchmarks, edging out strong frontier models such as GPT-4o and Claude 3.5 Sonnet.

See details on our paper at https://arxiv.org/abs/2410.01257 - as a preview, this model can correctly the question How many r in strawberry? without specialized prompting or additional reasoning tokens:

A sweet question!
Let’s count the “R”s in “strawberry”:
1. S
2. T
3. R
4. A
5. W
6. B
7. E
8. R
9. R
10. Y
There are **3 “R”s** in the word “strawberry”.

Terms of use

By accessing this model, you are agreeing to the LLama 3.1 terms and conditions of the license, acceptable use policy and Meta’s privacy policy

RewardBench Primary Dataset LeaderBoard

As of 1 Oct 2024, Llama-3.1-Nemotron-70B-Reward performs best Overall on RewardBench as well as with strong performance in Chat, Safety and Reasoning categories among the models below.

Model Type of Data Used For Training Overall Chat Chat-Hard Safety Reasoning
Llama-3.1-Nemotron-70B-Reward Permissive Licensed Data Only (CC-BY-4.0) 94.1 97.5 85.7 95.1 98.1
Skywork-Reward-Gemma-2-27B Includes GPT4 Generated Data 93.8 95.8 91.4 91.9 96.1
TextEval-Llama3.1-70B Not disclosed 93.5 94.1 90.1 93.2 96.4
Skywork-Critic-Llama-3.1-70B Not fully disclosed 93.3 96.6 87.9 93.1 95.5
SFR-LLaMa-3.1-70B-Judge-r Not fully disclosed 92.7 96.9 84.8 91.6 97.6
Nemotron-4-340B-Reward Permissive Licensed Data Only (CC-BY-4.0) 92.0 95.8 87.1 91.5 93.7
ArmoRM-Llama3-8B-v0.1 Includes GPT4 Generated Data 90.8 96.9 76.8 92.2 97.3
Cohere May 2024 Not disclosed 89.5 96.4 71.3 92.7 97.7
Llama3-70B-SteerLM-RM Permissive Licensed Data Only (CC-BY-4.0) 88.8 91.3 80.3 92.8 90.7
Google Gemini Pro 1.5 Not disclosed 88.1 92.3 80.6 87.5 92.0
GPT-4o-2024-08-06 Not disclosed 86.7 96.1 76.1 88.1 86.6
claude-3-5-sonnet-20240620 Not disclosed 84.2 96.4 74.0 81.6 84.7
Meta-Llama-3.1-70B-Instruct Not fully disclosed 84.0 97.2 70.2 82.8 86.0

To better understand why Llama-3.1-Nemotron-70B-Reward does less well in the Chat-Hard category, we analyze the scores for each consistutent subset under the Chat-Hard category. We find that on categories that uses human annotations as ground truth, Llama-3.1-Nemotron-70B-Reward performs similar to Skywork-Reward-Gemma-2-27B (<= 2.2% difference). On the other hand, when GPT-4 annotations are used as Ground-Truth, Llama-3.1-Nemotron-70B-Reward trails substantially behind Skywork-Reward-Gemma-2-27B (by 10.8 to 19.2%). This suggests that Skywork-Reward-Gemma-2-27B can better modelling GPT-4 preferences (but not human-annotated preferences), likely contributed by the inclusion of GPT-4 annotated training data used to train it found in the OffSetBias dataset as part of the Skywork-Reward-Preference-80k.

Model Type of Data Used For Training Chat-Hard LLMBar-Adversarial-Manual LLMBar-Adversarial-Neighbour LLMBar-Natural LLMBar-Adversarial-GPTInst LLMBar-Adversarial-GPTOut MT-Bench-Hard
Human as Ground Truth Human as Ground Truth Human as Ground Truth GPT-4 as Ground Truth GPT-4 as Ground Truth GPT-4 as Ground Truth
Llama-3.1-Nemotron-70B-Reward Permissive Licensed Data Only (CC-BY-4.0) 85.7 76.1 88.8 95.0 87.0 72.3 75.7
Skywork-Reward-Gemma-2-27B Includes GPT4 Generated Data 91.4 78.3 89.6 96.0 97.8 91.5 86.5

Usage:

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 v4.44.0, torch v2.4.0 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-3.1-Nemotron-70B-Reward-HF"

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

prompt = "What is 1+1?"
good_response = "1+1=2"
bad_response = "1+1=3"

for response in [good_response, bad_response]:
    messages = [{'role': "user", "content": prompt}, {'role': "assistant", "content": response}]
    tokenized_message = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=False, 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=1, return_dict_in_generate=True, output_scores=True)
    reward = response_token_ids['scores'][0][0][0].item()
    print(reward)

# reward for good_response = -3.28125
# reward for bad_response = -4.28125

Contact

E-Mail: Zhilin Wang

Citation

If you find this model useful, please cite the following works

@misc{wang2024helpsteer2preferencecomplementingratingspreferences,
      title={HelpSteer2-Preference: Complementing Ratings with Preferences}, 
      author={Zhilin Wang and Alexander Bukharin and Olivier Delalleau and Daniel Egert and Gerald Shen and Jiaqi Zeng and Oleksii Kuchaiev and Yi Dong},
      year={2024},
      eprint={2410.01257},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.01257}, 
}

@misc{wang2024helpsteer2,
      title={HelpSteer2: Open-source dataset for training top-performing reward models}, 
      author={Zhilin Wang and Yi Dong and Olivier Delalleau and Jiaqi Zeng and Gerald Shen and Daniel Egert and Jimmy J. Zhang and Makesh Narsimhan Sreedhar and Oleksii Kuchaiev},
      year={2024},
      eprint={2406.08673},
      archivePrefix={arXiv},
      primaryClass={id='cs.CL' full_name='Computation and Language' is_active=True alt_name='cmp-lg' in_archive='cs' is_general=False description='Covers natural language processing. Roughly includes material in ACM Subject Class I.2.7. Note that work on artificial languages (programming languages, logics, formal systems) that does not explicitly address natural-language issues broadly construed (natural-language processing, computational linguistics, speech, text retrieval, etc.) is not appropriate for this area.'}
}

References(s):

Model Architecture:

Architecture Type: Transformer
Network Architecture: Llama 3.1

Input:

Input Type(s): Text
Input Format: String
Input Parameters: NA
Other Properties Related to Input: Provided text must be within 4096 tokens

Output:

Output Type(s): Float
Output Format: One Single Float
Output Parameters: NA
Other Properties Related to Output: NA

Software Integration:

Supported Hardware Microarchitecture Compatibility:

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

Model Version:

v1.0

Training & Evaluation:

Datasets:

Data Collection Method by dataset

  • [Hybrid: Human, Synthetic]

Labeling Method by dataset

  • [Human]

Link:

Properties (Quantity, Dataset Descriptions, Sensor(s)):

  • Built to make more models more aligned with human preference - specifically more helpful, factually-correct, coherent, and customizable based on complexity and verbosity.

Inference:

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

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.

Please report security vulnerabilities or NVIDIA AI Concerns here.