|
--- |
|
license: llama3.1 |
|
library_name: nemo |
|
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 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. |
|
|
|
A HuggingFace Transformers compatible version converted from this model is available at [https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward-HF](https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward-HF) |
|
|
|
Try hosted inference for free at [build.nvidia.com](https://build.nvidia.com/nvidia/llama-3_1-nemotron-70b-reward) - 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](https://tatsu-lab.github.io/alpaca_eval/) of 57.6, [Arena Hard](https://github.com/lmarena/arena-hard-auto) of 85.0 and [GPT-4-Turbo MT-Bench](https://github.com/lm-sys/FastChat/pull/3158) of 8.98, which are known to be predictive of [LMSys Chatbot Arena Elo](https://huggingface.co/spaces/lmsys/chatbot-arena-leaderboard) This Instruct model is available at [Llama-3.1-Nemotron-70B-Instruct](https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct) as .nemo model and [Llama-3.1-Nemotron-70B-Instruct-HF](https://huggingface.co/nvidia/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](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](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/LICENSE), [acceptable use policy](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/USE_POLICY.md) and [Meta’s privacy policy](https://www.facebook.com/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](https://huggingface.co/datasets/NCSOFT/offsetbias) as part of the [Skywork-Reward-Preference-80k](https://huggingface.co/datasets/Skywork/Skywork-Reward-Preference-80K-v0.1). |
|
|
|
|
|
|
|
| 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 with [NeMo Aligner](https://github.com/NVIDIA/NeMo-Aligner) following [SteerLM training user guide](https://docs.nvidia.com/nemo-framework/user-guide/latest/modelalignment/steerlm.html). |
|
|
|
1. Spin up an inference server within the NeMo container (`docker pull nvcr.io/nvidia/nemo.24.05.llama3.1`) |
|
|
|
|
|
```python |
|
HF_HOME=<YOUR_HF_HOME_CONTAINING_TOKEN_WITH_LLAMA31_70B_ACCESS> \ |
|
python /opt/NeMo-Aligner/examples/nlp/gpt/serve_reward_model.py \ |
|
rm_model_file=Llama-3.1-Nemotron-70B-Reward \ |
|
trainer.num_nodes=1 \ |
|
trainer.devices=8 \ |
|
++model.tensor_model_parallel_size=8 \ |
|
++model.pipeline_model_parallel_size=1 \ |
|
inference.micro_batch_size=2 \ |
|
inference.port=1424 |
|
``` |
|
|
|
2. Annotate data files using the served reward model. As an example, this can be the Open Assistant train/val files. Then follow the next step to train a SteerLM model based on [SteerLM training user guide](https://docs.nvidia.com/nemo-framework/user-guide/latest/modelalignment/steerlm.html#step-5-train-the-attribute-conditioned-sft-model) . |
|
|
|
Please note that this script rounds the predicted floats to the nearest int (between 0 and 4 inclusive), as it's meant for SteerLM training. |
|
For other use cases (e.g. reward bench measurement, response filtering/ranking), we recommend using the floats directly, which can be done by commenting out [two lines of code in NeMo-Aligner](https://github.com/NVIDIA/NeMo-Aligner/blob/main/examples/nlp/data/steerlm/attribute_annotate.py#L135-136) |
|
|
|
```python |
|
python /opt/NeMo-Aligner/examples/nlp/data/steerlm/preprocess_openassistant_data.py --output_directory=data/oasst |
|
python /opt/NeMo-Aligner/examples/nlp/data/steerlm/attribute_annotate.py \ |
|
--input-file=data/oasst/train.jsonl \ |
|
--output-file=data/oasst/train_labeled.jsonl \ |
|
--port=1424 |
|
``` |
|
|
|
3. Alternatively, this can be any conversational data file (in .jsonl) in the following format, where each line looks like |
|
|
|
```json |
|
{ |
|
"conversations": [ |
|
{"value": <user_turn_1>, "from": "User", "label": None}, |
|
{"value": <assistant_turn_1>, "from": "Assistant", "label": <formatted_label_1>}, |
|
{"value": <user_turn_2>, "from": "User", "label": None}, |
|
{"value": <assistant_turn_2>, "from": "Assistant", "label": <formatted_label_2>}, |
|
], |
|
"mask": "User" |
|
} |
|
``` |
|
|
|
Ideally, each ```<formatted_label_n>``` refers to the ground truth label for the assistant turn but if they are not available, we can also use ```helpfulness:-1```. It must not be ```None``` or an empty string. |
|
|
|
|
|
|
|
## Contact |
|
|
|
E-Mail: [Zhilin Wang](mailto:zhilinw@nvidia.com) |
|
|
|
|
|
## Citation |
|
|
|
If you find this model useful, please cite the following works |
|
|
|
```bibtex |
|
@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): |
|
|
|
* [HelpSteer2-Preference](https://arxiv.org/abs/2410.01257) |
|
* [HelpSteer2](https://arxiv.org/abs/2406.08673) |
|
* [HelpSteer](https://arxiv.org/abs/2311.09528) |
|
* [SteerLM method](https://arxiv.org/abs/2310.05344) |
|
* [Introducing Llama 3.1: Our most capable models to date](https://ai.meta.com/blog/meta-llama-3-1/) |
|
* [Meta's Llama 3.1 Webpage](https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_1) |
|
* [Meta's Llama 3.1 Model Card](https://github.com/meta-llama/llama-models/blob/main/models/llama3_1/MODEL_CARD.md) |
|
|
|
|
|
## Model Architecture: |
|
**Architecture Type:** Transformer <br> |
|
**Network Architecture:** Llama 3.1 <br> |
|
|
|
## Input: |
|
**Input Type(s):** Text <br> |
|
**Input Format:** String <br> |
|
**Input Parameters:** NA <br> |
|
**Other Properties Related to Input:** Provided text must be within 4096 tokens <br> |
|
|
|
## Output: |
|
**Output Type(s):** Float <br> |
|
**Output Format:** One Single Float <br> |
|
**Output Parameters:** NA <br> |
|
**Other Properties Related to Output:** NA <br> |
|
|
|
## Software Integration: |
|
**Supported Hardware Microarchitecture Compatibility:** <br> |
|
* NVIDIA Ampere <br> |
|
* NVIDIA Hopper <br> |
|
* NVIDIA Turing <br> |
|
**Supported Operating System(s):** Linux <br> |
|
|
|
## Model Version: |
|
v1.0 |
|
|
|
# Training & Evaluation: |
|
|
|
## Datasets: |
|
|
|
**Data Collection Method by dataset** <br> |
|
* [Hybrid: Human, Synthetic] <br> |
|
|
|
**Labeling Method by dataset** <br> |
|
* [Human] <br> |
|
|
|
**Link:** |
|
* [HelpSteer2](https://huggingface.co/datasets/nvidia/HelpSteer2) |
|
|
|
**Properties (Quantity, Dataset Descriptions, Sensor(s)):** <br> |
|
* 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](https://developer.nvidia.com/triton-inference-server) <br> |
|
**Test Hardware:** H100, A100 80GB, A100 40GB <br> |
|
|
|
|
|
## 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](https://www.nvidia.com/en-us/support/submit-security-vulnerability/). |
|
|
|
Please report security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/). |
|
|
|
|
