DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models

🔹 This video covers DataFlex, a framework for dynamically optimizing training data in large language model training.
🔹 It unifies data selection, data mixture optimization, and sample reweighting under one system.
🔹 Built on top of LLaMA-Factory, it is designed as a drop-in framework for practical LLM training workflows.
🔹 We look at reported results across Mistral-7B, Llama-3.2-3B, and Qwen2.5 settings, including MMLU, perplexity, and runtime efficiency.
🔹 The paper shows why data-centric training may become a major direction for scaling and improving LLMs.

#DataFlex #LLMTraining #DataCentricAI #LLaMAFactory #MMLU #MachineLearning

Видео DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models канала CosmoX