OpenDCAI/DataFlex

[NeurIPS 2026] DataFlex: A Unified Benchmark and Evaluation Platform for Data-Centric Training of Large Language Models

What it solves

DataFlex addresses the difficulty of reproducing and implementing data-centric training strategies for Large Language Models (LLMs). It provides a unified framework to dynamically manage training data during the optimization process, replacing the need to manually integrate fragmented and often undocumented repositories for data selection, mixing, and reweighting.

How it works

Built as an extension of LLaMA-Factory, DataFlex integrates directly into the LLM training loop to intelligently schedule data. It implements three primary data-centric mechanisms:

  • Data Selection: Uses strategies like gradient-based (LESS, NICE), loss-based, or distribution-based (NEAR, TSDS) methods to pick specific samples (e.g., "hard" samples) for training.
  • Data Mixture: Dynamically adjusts the ratio of data from different domains during training using methods like DoReMi (offline) or ODM (online).
  • Data Reweighting: Adjusts the weight of individual samples during backpropagation to emphasize preferred data using loss-based or batch-aware techniques.

Who it’s for

It is designed for AI researchers and developers who want to improve model performance and experimental reproducibility by applying advanced data-centric training techniques without building the infrastructure from scratch.

Highlights

  • LLaMA-Factory Integration: Acts as a drop-in replacement that is fully compatible with LLaMA-Factory.
  • Unified Implementation: Consolidates multiple data-centric algorithms (LESS, NICE, DoReMi, ODM, etc.) into one platform.
  • Scalability: Supports gradient computation under DeepSpeed ZeRO-3 for training larger models.
  • Extensible: Includes a registry system and base class interfaces for adding new selectors, mixers, or weighters.

Related

  • 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 modes and agentic workflows.

  • Project
    OpenRLHF/OpenRLHF

    A high-performance RLHF framework that combines Ray and vLLM to scale reinforcement learning from human feedback for LLMs and VLMs.

  • Project
    axolotl-ai-cloud/axolotl

    A free and open-source LLM fine-tuning framework that streamlines post-training for text, vision, and audio models using a single YAML configuration.

  • Project
    ServiceNow/PipelineRL

    A scalable asynchronous reinforcement learning framework for LLM agents that uses in-flight weight updates to maximize GPU utilization while maintaining on-policy data freshness.

  • 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.