CYCLE-INSTRUCT: Fully Seed-Free Instruction Tuning via Dual Self-Training and Cycle Consistency

2025-08-22

AI Summary

Existing instruction tuning relies on manually annotated seed data or strong teacher models, while instruction back-translation remains constrained by the initial seed set, leading to error accumulation and inefficient utilization of unlabeled corpora. Method: We propose the first fully seed-free instruction tuning framework, built upon a dual-model architecture—answer generation and question generation—that establishes a bidirectional self-training loop. This loop leverages pseudo-label reconstruction, mutual supervision, and cycle-consistency constraints to automatically synthesize high-quality pseudo-instruction data from raw text alone. Contribution/Results: Our method eliminates dependence on human annotation and external models, avoids initialization bias, and substantially improves raw corpus utilization. On four benchmark task categories, it outperforms seed-based back-translation approaches and matches strong supervised baselines—demonstrating, for the first time, the feasibility and effectiveness of fully self-guided instruction tuning.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Unsupervised & Self-Supervised Learning
• Search and Optimization: Learning to Search

Application Category

• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling
• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Instruction tuning is vital for aligning large language models (LLMs) with human intent, but current methods typically rely on costly human-annotated seed data or powerful external teacher models. While instruction back-translation techniques reduce this dependency, they remain fundamentally tethered to an initial seed set, which limits full automation, introduces biases, and can lead to inefficient use of unlabeled corpora. In this paper, we propose Cycle-Instruct, a novel framework that achieves fully seed-free instruction tuning. Inspired by cycle consistency, Cycle-Instruct employs a dual self-training loop where two models-an answer generator and a question generator-are bootstrapped solely from raw, unlabeled text. These models mutually supervise each other by reconstructing original text segments from their counterpart's generated pseudo-labels, effectively learning from the intrinsic structure of the data without any human-provided seeds. We demonstrate Cycle-Instruct's efficacy across four diverse data tracks, including general instruction-following, domain-specific tasks, dialogue logs, and plain text. Our extensive experiments show that Cycle-Instruct not only outperforms seed-driven back-translation baselines but also achieves performance comparable to strongly supervised methods.

Problem

• Achieves fully seed-free instruction tuning without human data
• Eliminates dependency on initial seed sets to avoid biases
• Uses dual self-training to learn from unlabeled text corpora

Innovation

• Seed-free dual self-training loop
• Cycle consistency for mutual supervision
• Bootstrapping from raw unlabeled text