open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_SimNPO_lr2e-05_b3.5_a1_d1_g0.25_ep10
The open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_SimNPO_lr2e-05_b3.5_a1_d1_g0.25_ep10 model is a 1 billion parameter instruction-tuned language model, likely based on the Llama 3.2 architecture. This model is specifically designed for 'unlearning' tasks, indicated by 'unlearn_tofu' and 'forget10', suggesting it has been modified to remove specific information or behaviors. Its primary differentiation lies in its application of SimNPO (Simple Negative Preference Optimization) for targeted unlearning, making it suitable for research and development in model safety and content moderation.
Loading preview...
Model Overview
This model, open-unlearning/unlearn_tofu_Llama-3.2-1B-Instruct_forget10_SimNPO_lr2e-05_b3.5_a1_d1_g0.25_ep10, is a 1 billion parameter instruction-tuned language model. It is derived from the Llama 3.2 architecture and has undergone a specialized training process focused on 'unlearning' specific information or behaviors, as indicated by the 'unlearn_tofu' and 'forget10' components in its name. The model utilizes SimNPO (Simple Negative Preference Optimization) with specific hyperparameters (learning rate 2e-05, batch size 3.5, alpha 1, delta 1, gamma 0.25, 10 epochs) to achieve this targeted unlearning.
Key Characteristics
- Parameter Count: 1 billion parameters, offering a balance between performance and computational efficiency.
- Context Length: Supports a substantial context length of 32768 tokens, enabling processing of longer inputs.
- Unlearning Focus: Explicitly trained for unlearning tasks, making it distinct from general-purpose instruction-tuned models.
- SimNPO Method: Employs Simple Negative Preference Optimization, a technique designed to modify model behavior by reducing preference for certain outputs or knowledge.
Potential Use Cases
- Research in Model Unlearning: Ideal for exploring techniques to remove unwanted information or biases from large language models.
- Safety and Alignment Studies: Can be used to investigate methods for improving model safety and aligning AI behavior with ethical guidelines.
- Controlled Content Generation: Potentially useful in scenarios requiring models to avoid generating specific types of content or discussing particular topics.