Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Position: Federated Foundation Language Model Post-Training Should Focus on Open-Source Models

Nikita Agrawal, Simon Mertel, Ruben Mayer

TL;DR

<3-5 sentence high-level summary>This position paper argues that federated foundation language model post-training should prioritize open-source (and open-weight) models to preserve privacy, autonomy, and transparency in FL. It defines a four-way openness taxonomy and analyzes how openness enables or restricts post-training methods (FFT, LoRA, adapters, prompt/instruction tuning, RLHF) while examining licensing and data considerations. The authors contend that open models align with FL principles and present a privacy/security analysis and a model-selection guide, while outlining significant risks associated with closed/black-box models. They conclude that a disciplined focus on open models yields more trustworthy, regulatory-compatible, and controllable FL post-training outcomes.

Abstract

Post-training of foundation language models has emerged as a promising research domain in federated learning (FL) with the goal to enable privacy-preserving model improvements and adaptations to user's downstream tasks. Recent advances in this area adopt centralized post-training approaches that build upon black-box foundation language models where there is no access to model weights and architecture details. Although the use of black-box models has been successful in centralized post-training, their blind replication in FL raises several concerns. Our position is that using black-box models in FL contradicts the core principles of federation such as data privacy and autonomy. In this position paper, we critically analyze the usage of black-box models in federated post-training, and provide a detailed account of various aspects of openness and their implications for FL.

Position: Federated Foundation Language Model Post-Training Should Focus on Open-Source Models

TL;DR

<3-5 sentence high-level summary>This position paper argues that federated foundation language model post-training should prioritize open-source (and open-weight) models to preserve privacy, autonomy, and transparency in FL. It defines a four-way openness taxonomy and analyzes how openness enables or restricts post-training methods (FFT, LoRA, adapters, prompt/instruction tuning, RLHF) while examining licensing and data considerations. The authors contend that open models align with FL principles and present a privacy/security analysis and a model-selection guide, while outlining significant risks associated with closed/black-box models. They conclude that a disciplined focus on open models yields more trustworthy, regulatory-compatible, and controllable FL post-training outcomes.

Abstract

Post-training of foundation language models has emerged as a promising research domain in federated learning (FL) with the goal to enable privacy-preserving model improvements and adaptations to user's downstream tasks. Recent advances in this area adopt centralized post-training approaches that build upon black-box foundation language models where there is no access to model weights and architecture details. Although the use of black-box models has been successful in centralized post-training, their blind replication in FL raises several concerns. Our position is that using black-box models in FL contradicts the core principles of federation such as data privacy and autonomy. In this position paper, we critically analyze the usage of black-box models in federated post-training, and provide a detailed account of various aspects of openness and their implications for FL.

Paper Structure

This paper contains 30 sections, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Openness, Post-Training Methods and Federated Learning
  3. Model Openness vis-à-vis Post-Training Methods
  4. Degree of Openness
  5. Post-Training Methods
  6. Licensing
  7. Federated Learning on Foundation Models
  8. Federated Black-Box Post-Training Methods
  9. Open Models Fit Federated Learning Better
  10. Alignment with Federated Learning Principles
  11. Weights and code
  12. Licenses
  13. Data
  14. Closed Models Bring Many Risks
  15. Privacy and Security
  16. ...and 15 more sections

Figures (2)

  • Figure 1: Federated post-training of gray-box models. Top panel: Gray-box LLM, where the base model parameters are frozen and inaccessible, while post-training is enabled through APIs allowing soft prompts or lightweight adapter layers. Bottom panel: Two FL instantiations under the gray-box setting: (a) clients collaboratively learn soft prompts via local training and share only prompt models updates with a central aggregator, and (b) clients train local adapter layers attached to the frozen LLM and share only adapter parameters. In both cases, the base model remains frozen and proprietary, while the provider exposes only necessary trainable layers.
  • Figure 2: Decision tree for selecting foundation model type based on privacy, security, autonomy, and heterogeneity