thrnn/qwen2.5-1.5b-sft-dare-resta
thrnn/qwen2.5-1.5b-sft-dare-resta is a 1.5 billion parameter language model based on the Qwen2.5 architecture, created by thrnn. This model is a merge using the Task Arithmetic method, combining Qwen/Qwen2.5-1.5B-Instruct with a 'harmful_lora' component, using thrnn/qwen2.5-1.5b-medical-sft-dare as its base. Its unique merging strategy suggests a focus on modifying or exploring specific behavioral characteristics of the base model, potentially for safety research or content moderation. The model has a context length of 32768 tokens.
Loading preview...
Model Overview
thrnn/qwen2.5-1.5b-sft-dare-resta is a 1.5 billion parameter language model developed by thrnn, built upon the Qwen2.5 architecture. It leverages a unique merging approach known as Task Arithmetic to combine different pre-trained components. The model's base is thrnn/qwen2.5-1.5b-medical-sft-dare, which has been merged with Qwen/Qwen2.5-1.5B-Instruct and an additional component referred to as outputs/model_harmful_lora.
Merge Details
The core of this model's creation lies in its specific merge configuration. The Qwen/Qwen2.5-1.5B-Instruct model was integrated with a weight of 1.0, while the outputs/model_harmful_lora component was applied with a negative weight of -1.0. This inverse weighting for the 'harmful_lora' suggests an experimental approach, possibly aimed at mitigating or analyzing certain undesirable behaviors or characteristics within the base model. The merge was performed using mergekit and float16 precision.
Potential Use Cases
Given its unique merging strategy, this model could be particularly interesting for:
- Behavioral Research: Investigating the impact of specific LORA components on model outputs, especially concerning safety or harmful content generation.
- Model Debiasing Experiments: Exploring methods to reduce or modify pre-existing biases or undesirable traits in instruction-tuned models.
- Advanced Merging Techniques: As a case study for applying Task Arithmetic with negative weights to achieve specific model modifications.
Developers interested in the effects of model merging on specific behavioral aspects, rather than general performance improvements, may find this model valuable for their research.