OpenDCAI/DataFlex

Data-centric LLM training with dynamic sample selection, domain mixture optimization, and example reweighting inside the LLaMA-Factory training loop.

What it solves

DataFlex addresses the difficulty of managing training data during the LLM optimization process. Instead of using a static dataset, it allows developers to dynamically adjust which samples are used, how they are mixed, and how much weight they contribute to the model's learning process, which often leads to better model performance and experimental reproducibility.

How it works

Built as a dynamic training framework on top of LLaMA-Factory, DataFlex integrates several data-centric training strategies into a unified system. It provides reproducible implementations of three core capabilities:

  • Data Selection: Dynamically chooses training samples based on strategies like gradient-based (LESS, NICE) or loss-based methods, focusing on "hard" samples.
  • Data Mixture: Adjusts the ratio of data from different domains (e.g., using DOREMI or ODM) during training to optimize the balance between different types of information.
  • Data Reweighting: Modifies the sample weights during backpropagation to emphasize specific data points preferred by the model.

Who it’s for

Researchers and developers training Large Language Models (LLMs) who want more flexible and powerful control over their training data scheduling and optimization.

Highlights

  • LLaMA-Factory Integration: Acts as a drop-in replacement for LLaMA-Factory, maintaining full compatibility.
  • Unified Framework: Consolidates multiple difficult-to-reproduce research repositories into one place.
  • Scalability: Supports gradient computation under DeepSpeed ZeRO-3 for training larger-scale models.
  • Data-Centric Approach: Focuses on the data scheduling during the training loop rather than just pre-processing.

Related

  • Project
    OpenRLHF/OpenRLHF

    A high-performance RLHF framework using Ray and vLLM to scale reinforcement learning for LLMs and VLMs, featuring a unified agent-based design for single and multi-turn training.

  • Project
    zjunlp/LightMem

    A lightweight memory management framework for LLMs and AI agents that provides efficient storage, retrieval, and update mechanisms for long-term memory.

  • Project
    agentscope-ai/Trinity-RFT

    A general-purpose framework for reinforcement fine-tuning of LLMs that decouples exploration, training, and data buffering to support diverse RL algorithms and agentic workflows.

  • Project
    OpenDCAI/DataFlow

    DataFlow is an open‑source Python framework that lets you build, run, and share low‑code pipelines for generating, cleaning, and evaluating LLM training data. It provides a library of reusable operators, a visual WebUI, an AI‑assistant for auto‑creating pipelines, and a Ray‑based distributed execution layer.

  • Project
    EvolvingLMMs-Lab/lmms-engine

    A unified, high-performance training engine for multimodal models that integrates distributed training, kernel fusion, and advanced optimizers to enable scalable pretraining and fine-tuning.