ssadqdsacf/cross-unlearning-case6-qwen35-4b-ga-difference-epoch10
The ssadqdsacf/cross-unlearning-case6-qwen35-4b-ga-difference-epoch10 is a 4.5 billion parameter language model based on the Qwen 3.5 architecture, specifically initialized from an epoch-10 SFT model. This model focuses on "ga_difference unlearning," a technique aimed at removing specific information or behaviors. It is one of six formal matrix models, providing a directly loadable merged model and a LoRA adapter, and is designed for research into unlearning methodologies.
Loading preview...
Model Overview
The ssadqdsacf/cross-unlearning-case6-qwen35-4b-ga-difference-epoch10 is a 4.5 billion parameter language model derived from the Qwen 3.5 architecture. It represents a specific iteration within a broader research matrix, initialized from an epoch-10 Supervised Fine-Tuning (SFT) model. The primary focus of this model is "ga_difference unlearning," a method for selectively removing learned information or behaviors from the model.
Key Capabilities & Characteristics
- Unlearning Research: This model is specifically designed for investigating and implementing "ga_difference unlearning" techniques.
- Provenance Tracking: Detailed machine-readable provenance is provided, including protocol, SHA256 fingerprints, SFT selection details, and evaluation methodology.
- Flexible Deployment: The repository offers both a directly loadable merged model and a corresponding LoRA adapter for adaptable integration.
- Evaluation Insights: Includes verified three-direction classification, mask, and generation summaries, utilizing fixed 4-shot greedy decoding for evaluation.
Good For
- Research into Model Unlearning: Ideal for academics and researchers exploring methods to remove specific data or biases from large language models.
- Comparative Studies: As one of six formal matrix models, it serves as a baseline for comparing different unlearning approaches or model states.
- Understanding Model Evolution: The detailed provenance allows for in-depth analysis of the model's training and selection process at a specific epoch.