| Documentation | Blog |

ArcticTraining: Simplifying and Accelerating Post-Training for LLMs

ArcticTraining is a framework designed to simplify and accelerate the post-training process for large language models (LLMs). It addresses challenges in current frameworks, such as limited support for rapid prototyping and the lack of native data generation tools, by offering modular trainer designs, simplified code structures, and integrated pipelines for creating and cleaning synthetic data. These features enable users to enhance LLM capabilities, like code generation and complex reasoning, with greater efficiency and flexibility. Read more about ArcticTraining in our blog.

Quickstart

To get started training a model with ArcticTraining, follow the steps below:

1. Install the ArcticTraining package and its dependencies:

pip install arctic-training

2. Create a training recipe YAML that uses the built-in Supervised Fine-Tuning (SFT) trainer:

type: sft
micro_batch_size: 2
model:
  name_or_path: meta-llama/Meta-Llama-3.1-8B-Instruct
data:
  sources:
    - HuggingFaceH4/ultrachat_200k
checkpoint:
  - type: huggingface
    save_end_of_training: true
    output_dir: ./fine-tuned-model

3. Run the training recipe with the ArcticTraining CLI (see below). This will use the DeepSpeed launcher behind the scenes, you can pass any compatible DeepSpeed launcher arguments to the ArcticTraining CLI (e.g., --num_nodes, --num_gpus).

arctic_training path/to/sft-recipe.yaml

Projects

The projects folder contains all special projects we release that build on-top of ArcticTraining. For example yamls and to dive deeper into the training code please see the following projects:

• SwiftKV
• Speculative Decoding

Customize Training

To customize the training workflow, you can modify the training recipe YAML we created in step 3 above. For example, you can change the model, dataset, checkpoint, or other settings to meet your specific requirements. A full list of configuration options can be found on the configuration documentation page.

Creating a New Trainer

If you want to create a new trainer, you can do so by subclassing the Trainer or SFTTrainer classes and implementing the necessary modifications. For example, you could create a new trainer from SFTTrainer that uses a different loss function:

from arctic_training import register
from arctic_training import SFTTrainer

@register
class CustomTrainer(SFTTrainer):
   name = "my_custom_trainer"

   def loss(self, batch):
       # Custom loss function implementation
       return loss

Remember to register this new trainer using the @register decorator so that it can be used in training recipes. By default, ArcticTraining looks for a train.py in the current working directory to find custom trainers. You can also specify a custom path to the trainers with the code field in your training recipe:

type: my_custom_trainer
code: path/to/custom_trainers.py
model:
 name_or_path: meta-llama/Meta-Llama-3.1-8B-Instruct
data:
 sources:
   - HuggingFaceH4/ultrachat_200k