Mic-Fundraiser/fundraising-assistant
Mic-Fundraiser/fundraising-assistant is a 3.1 billion parameter conversational model developed by Mic-Fundraiser, fine-tuned from Qwen/Qwen2.5-3B-Instruct. Specialized in fundraising and non-profit fund collection, it is optimized for generating direct and competent Italian text for donation appeals, letters, and proposals. This model excels at adopting a "young professor" style, avoiding unnecessary preambles or lengthy lists, and is suitable for rapid inference on CPU/light GPU.
Loading preview...
Overview
Mic-Fundraiser/fundraising-assistant is a 3.1 billion parameter conversational model, fine-tuned from Qwen/Qwen2.5-3B-Instruct. It is specifically designed for fundraising and non-profit fund collection in Italian.
Key Capabilities
- Specialized Content Generation: Optimized for writing direct mail (DEM), letters, and proposals for non-profit organizations.
- Distinctive Style: Generates text with a "competent, direct, young professor" tone, avoiding unnecessary preambles, antitheses, or endless lists.
- Language: Fully specialized in Italian.
Training Details
The model underwent fine-tuning (SFT + LoRA, then merge) using the Mic-Fundraiser/fundraising-italian-sft dataset, which consists of 1273 examples in OpenAI messages format. Training involved QLoRA 4-bit on the base Qwen2.5-3B-Instruct, with specific LoRA parameters (r=16, alpha=32, dropout=0.05, target=attn+MLP) over 3 epochs with a 2e-4 learning rate and cosine schedule.
Limitations
As a 3B parameter model, it is efficient for rapid inference on CPU/light GPU. However, for complex reasoning tasks or very long contexts, it may perform less effectively than larger models. The relatively small dataset (1273 examples) means the fine-tuning primarily imparts a specific style rather than new factual knowledge.