WisdomShell/ADG-Alpaca-GPT4-LLaMa3-8B is a model developed by Bo Li, Mingda Wang, Shikun Zhang, and Wei Ye, demonstrating the effectiveness of the Answer Divergence-Guided Selection (ADG) method for instruction data selection. ADG improves instruction tuning by scoring instructions based on the geometric structure of multiple sampled answers, rather than a single reference. This approach consistently enhances performance across reasoning, knowledge, and coding benchmarks under a fixed data budget. The repository provides the pipeline for multi-sample answer generation, instruction embedding, ADG scoring, and model training.
Loading preview...
Overview of ADG-Alpaca-GPT4-LLaMa3-8B
This model is a demonstration of the Answer Divergence-Guided Selection (ADG) method, a novel approach to instruction data selection for improving instruction tuning. Developed by Bo Li, Mingda Wang, Shikun Zhang, and Wei Ye, ADG addresses the challenge of selecting high-quality examples under a fixed data budget by analyzing the geometric structure of multiple answers sampled from a base model under stochastic decoding.
Key Capabilities and Methodology
- Geometry-aware Scoring: ADG scores instructions by sampling multiple answers, mapping them into a representation space, and computing geometry-aware scores based on dispersion magnitude and shape anisotropy.
- Bin-wise Selection: To ensure semantic coverage, ADG performs proportional selection within semantic bins after ranking examples by their combined scores.
- Improved Instruction Tuning: The method consistently improves instruction tuning quality across various backbones (e.g., LLaMA, Qwen), public instruction pools (e.g., Alpaca-GPT4, WizardLM, CoT), and benchmarks covering reasoning, knowledge, and coding tasks.
- Comprehensive Pipeline: The repository provides a practical pipeline for multi-sample answer generation, instruction embedding and clustering, ADG scoring and subset selection, model training, and benchmark evaluation.
Good for
- Researchers and developers focused on instruction data selection and improving instruction tuning efficiency.
- Experimenting with methods to enhance model performance under limited data budgets.
- Understanding and implementing advanced techniques for data curation in LLM training.