DataFlex

🎉 If you like our project, please give us a star ⭐ on GitHub for the latest update.

简体中文 | English

DataFlex is an advanced dynamic training framework built on top of LlamaFactory.
It intelligently schedules data during training, supporting dynamic sample selection, domain ratio adjustment, and dynamic weighting, aiming to improve both training efficiency and final model performance.

DataFlex integrates seamlessly with LlamaFactory, offering researchers and developers more flexible and powerful training control.

---

✨ Features

• Dynamic Select Trainer: Dynamically selects training samples according to a given strategy (e.g., focus on “hard” samples).
• Dynamic Mix Trainer: Dynamically adjusts the ratio of data from different domains during training.
• Dynamic Weight Trainer: Dynamically adjusts sample weights during backpropagation to emphasize data preferred by the model.
• Full compatibility with LlamaFactory, drop-in replacement.

---

🚀 Installation

git clone https://github.com/OpenDCAI/DataFlex-Preview.git
cd DataFlex-Preview
pip install -e .
pip install llamafactory

---

📌 Usage Example

The launch command is similar to LlamaFactory. Below is an example using LESS :

FORCE_TORCHRUN=1 DISABLE_VERSION_CHECK=1 dataflex-cli train examples/train_lora/less.yaml

Unlike vanilla LlamaFactory, your .yaml config file must also include DataFlex-specific parameters.

---

🔑 Select Trainer Configuration

# Select Trainer parameters
train_type: dynamic_select    # [dynamic_select, dynamic_mix, dynamic_weighting, static]
component_name: loss          # selection strategy, e.g. loss / less
components_cfg_file: src/dataflex/configs/components.yaml
warmup_step: 200              # warmup steps before the first selection
update_step: 1000             # trigger selection every N steps
update_times: 2               # number of times selection will be performed

Parameter Details:

• train_type: Training mode (dynamic select / dynamic mix / dynamic weighting / static).
• component_name: Selector strategy (e.g., loss or less).
• components_cfg_file: Path to the selector’s config file.
• warmup_step: Warmup steps before the first selection.
• update_step: Frequency of selector activation.
• update_times: Total number of times the selector will run.

---

📚 Core Concepts

• Trainer: Defines the training flow (selection, mixing, weighting).
• Selector: Encapsulates sample selection strategies.
• Components: Modular parameter definitions in config files.

---

🤝 Contributing

We welcome contributions of new trainers and selectors! Please ensure code formatting is consistent with the existing style before submitting a PR.