- We released Nemotron-4-340B Base, Instruct, Reward. The Instruct and Reward variants are trained in Nemo-Aligner. Please see the Helpsteer2 paper for more details on the reward model training.
- We are excited to announce the release of accelerated generation support in our RLHF pipeline using TensorRT-LLM. For more information, please refer to our RLHF documentation.
- NeMo-Aligner Paper is now out on arxiv!
NeMo-Aligner is a scalable toolkit for efficient model alignment. The toolkit has support for state-of-the-art model alignment algorithms such as SteerLM, DPO, and Reinforcement Learning from Human Feedback (RLHF). These algorithms enable users to align language models to be more safe, harmless, and helpful. Users can perform end-to-end model alignment on a wide range of model sizes and take advantage of all the parallelism techniques to ensure their model alignment is done in a performant and resource-efficient manner. For more technical details, please refer to our paper.
The NeMo-Aligner toolkit is built using the NeMo Framework, which enables scalable training across thousands of GPUs using tensor, data, and pipeline parallelism for all alignment components. Additionally, our checkpoints are cross-compatible with the NeMo ecosystem, facilitating inference deployment and further customization (https://github.com/NVIDIA/NeMo-Aligner).
The toolkit is currently in it's early stages. We are committed to improving the toolkit to make it easier for developers to pick and choose different alignment algorithms to build safe, helpful, and reliable models.
- SteerLM: Attribute Conditioned SFT as an (User-Steerable) alternative to RLHF.
- Llama3-70B-SteerLM-Chat aligned with NeMo-Aligner.
- Corresponding reward model Llama3-70B-SteerLM-RM.
- Learn more at our SteerLM and HelpSteer2 papers.
- Supervised Fine Tuning
- Reward Model Training
- Reinforcement Learning from Human Feedback using the PPO Algorithm
- Llama3-70B-PPO-Chat aligned with NeMo-Aligner using TRT-LLM.
- Direct Preference Optimization as described in Direct Preference Optimization: Your Language Model is Secretly a Reward Model
- Llama3-70B-DPO-Chat aligned with NeMo Aligner.
- Self-Play Fine-Tuning (SPIN) as described in Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models
For the latest stable release, please see the releases page. All releases come with a pre-built container. Changes within each release will be documented in CHANGELOG.
NeMo-Aligner has the same requirements as the NeMo Toolkit Requirements with the addition of PyTriton.
NeMo Aligner comes included with NeMo containers. On a machine with NVIDIA GPUs and drivers installed run NeMo container:
docker run --gpus all -it --rm --shm-size=8g --ulimit memlock=-1 --ulimit stack=67108864 nvcr.io/nvidia/nemo:24.07Once you are inside the container, NeMo-Aligner is already installed and together with NeMo and other tools can be found under /opt/ folder.
Please follow the same steps as outlined in the NeMo Toolkit Installation Guide. After installing NeMo, execute the following additional command:
pip install nemo-alignerAlternatively, if you prefer to install the latest commit:
pip install .We provide an official NeMo-Aligner Dockerfile which is based on stable, tested versions of NeMo, Megatron-LM, and TransformerEngine. The primary objective of this Dockerfile is to ensure stability, although it might not always reflect the very latest versions of those three packages. You can access our Dockerfile here.
Alternatively, you can build the NeMo Dockerfile here NeMo Dockerfile and add RUN pip install nemo-aligner at the end.
- We will continue improving the stability of the PPO learning phase.
- Improve the performance of RLHF.
- Add TRT-LLM inference support for Rejection Sampling.
We welcome community contributions! Please refer to CONTRIBUTING.md for guidelines.
@misc{shen2024nemoaligner,
title={NeMo-Aligner: Scalable Toolkit for Efficient Model Alignment},
author={Gerald Shen and Zhilin Wang and Olivier Delalleau and Jiaqi Zeng and Yi Dong and Daniel Egert and Shengyang Sun and Jimmy Zhang and Sahil Jain and Ali Taghibakhshi and Markel Sanz Ausin and Ashwath Aithal and Oleksii Kuchaiev},
year={2024},
eprint={2405.01481},
archivePrefix={arXiv},
primaryClass={cs.CL}
}
This toolkit is licensed under the Apache License, Version 2.0.