thrnn/qwen2.5-1.5b-sft-resta
thrnn/qwen2.5-1.5b-sft-resta is a 1.5 billion parameter language model based on the Qwen2.5 architecture, created by thrnn through a Task Arithmetic merge. This model integrates Qwen/Qwen2.5-1.5B-Instruct with a harmful LORA model, using thrnn/qwen2.5-1.5b-medical-sft-lora as its base. It is designed for specific applications requiring a modified response profile from its base models, leveraging its 32768 token context length.
Loading preview...
Model Overview
thrnn/qwen2.5-1.5b-sft-resta is a 1.5 billion parameter language model built upon the Qwen2.5 architecture, developed by thrnn. This model was created using the Task Arithmetic merge method, a technique that combines the weights of different pre-trained models to achieve specific behavioral modifications.
Merge Details
The core of this model is a merge of two components:
- Qwen/Qwen2.5-1.5B-Instruct: A foundational instruction-tuned model from the Qwen series.
- outputs/model_harmful_lora: A LORA (Low-Rank Adaptation) model, likely intended to modify or counteract specific behaviors, particularly those related to harmful content.
The merging process utilized thrnn/qwen2.5-1.5b-medical-sft-lora as the base model, suggesting an intent to adjust its characteristics. The configuration applied a weight of 1.0 to the Qwen2.5-1.5B-Instruct component and a weight of -1.0 to the outputs/model_harmful_lora, indicating a subtractive or counteractive influence from the harmful LORA model.
Key Characteristics
- Architecture: Qwen2.5
- Parameter Count: 1.5 billion
- Context Length: 32768 tokens
- Merge Method: Task Arithmetic, allowing for fine-grained control over model behavior by combining specific model components.
Potential Use Cases
This model is particularly suited for applications where developers need to:
- Experiment with model behavior modification: Ideal for researchers exploring the effects of merging specific LORA adapters.
- Develop custom instruction-following models: Provides a base that has been specifically adjusted through the Task Arithmetic method.
- Address specific content generation requirements: The inclusion and negative weighting of a "harmful LORA" suggest an application in content moderation or safety research.