Model Overview

PrasannaPaithankar/qwen2.5-1.5b-sft-resta is a 1.5 billion parameter language model built upon the Qwen2.5 architecture, featuring a substantial 32K context length. This model was created using the Task Arithmetic merge method, a technique that combines pre-trained language models by adjusting their weights based on specific tasks.

Merge Details

The model's unique configuration involves merging two primary components:

• Qwen/Qwen2.5-1.5B-Instruct: This serves as a foundational instruction-tuned model.
• outputs/model_harmful_lora: A LoRA (Low-Rank Adaptation) component, which is integrated with a negative weight.

The merging process utilized thrnn/qwen2.5-1.5b-medical-sft-lora as the base model. The application of a negative weight to the 'harmful_lora' component suggests an intent to mitigate or modify specific behavioral characteristics introduced by that LoRA, potentially aiming for a more refined or safer output profile.

Key Characteristics

• Architecture: Qwen2.5-1.5B
• Parameter Count: 1.5 billion
• Context Length: 32,768 tokens
• Merge Method: Task Arithmetic, allowing for precise control over the influence of merged components.

Potential Use Cases

This model could be particularly useful for developers looking to:

• Experiment with Task Arithmetic for fine-grained control over model behavior.
• Explore methods for modifying or steering model outputs by subtracting specific LoRA influences.
• Develop applications requiring a compact yet capable Qwen2.5-based model with adjusted response characteristics.