PrasannaPaithankar/qwen2.5-1.5b-sft-dare-resta

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Apr 3, 2026Architecture:Transformer Warm

PrasannaPaithankar/qwen2.5-1.5b-sft-dare-resta is a 1.5 billion parameter language model based on the Qwen2.5 architecture, created by PrasannaPaithankar. This model is a merge of pre-trained language models using the Task Arithmetic method, specifically designed to modify the behavior of the base model. It integrates Qwen/Qwen2.5-1.5B-Instruct with a 'harmful_lora' component, suggesting a focus on adjusting or mitigating specific content generation characteristics.

Loading preview...

Model Overview

PrasannaPaithankar/qwen2.5-1.5b-sft-dare-resta is a 1.5 billion parameter language model derived from the Qwen2.5 architecture. It was created by PrasannaPaithankar using the mergekit tool, specifically employing the Task Arithmetic merge method.

Key Characteristics

  • Base Model: The merge uses PrasannaPaithankar/qwen2.5-1.5b-medical-sft-dare as its foundation.
  • Merged Components: It combines Qwen/Qwen2.5-1.5B-Instruct with a component identified as outputs/model_harmful_lora.
  • Merge Method: The Task Arithmetic method was applied, with Qwen/Qwen2.5-1.5B-Instruct having a weight of 1.0 and outputs/model_harmful_lora having a negative weight of -1.0. This negative weighting in Task Arithmetic typically aims to subtract or reduce specific learned behaviors or characteristics introduced by the negatively weighted model.

Potential Use Cases

This model is likely intended for use cases where fine-grained control over content generation is desired, particularly in modifying or 'restoring' the behavior of a base model by subtracting specific learned traits. The inclusion of a 'harmful_lora' with a negative weight suggests an effort to mitigate or remove undesirable outputs, potentially related to safety or specific content biases, from the Qwen2.5-1.5B-Instruct base.