tsilva/qwen2.5-3b-trump-style-merged-v2
The tsilva/qwen2.5-3b-trump-style-merged-v2 is a 3.1 billion parameter Qwen2.5-based language model, fine-tuned by tsilva to generate text in a Trump-like public speaking style. This merged 16-bit model is specifically designed for style-following chat experiments and evaluation, offering a distinct stylistic output. It is optimized for direct Transformer loading and is part of a versioned style-tuning release.
Loading preview...
Overview
This model, tsilva/qwen2.5-3b-trump-style-merged-v2, is a 3.1 billion parameter Qwen2.5-based language model developed by tsilva. It has been specifically fine-tuned to emulate a Trump-like public speaking style, making it unique among general-purpose LLMs. This particular artifact is a merged 16-bit model, generated through an llmstyler Runbook training job, and is part of a versioned style-tuning release.
Key Characteristics
- Base Model: Built upon
unsloth/Qwen2.5-3B-Instruct-bnb-4bit. - Style Tuning: Fine-tuned using the
tsilva/stylemix_trump-v2dataset to adopt a "trump_like_public_speaking" style. - Training Details: Trained for 5 epochs with a max sequence length of 2048, achieving a train loss of 1.323 and an eval loss of 1.369.
- Artifacts: This repository contains the merged 16-bit model, with associated QLoRA adapter, GGUF, and ONNX exports available in separate repositories.
Intended Use Cases
- Style-Following Experiments: Ideal for research and development focused on generating text in a specific, recognizable public speaking style.
- Evaluation: Suitable for evaluating the strength and consistency of style application in generated content.
- Direct Transformer Loading: Designed for straightforward integration into workflows that utilize direct Transformer model loading.
Limitations
Users should be aware that the model may sometimes over-apply the target style, potentially missing factual nuance or reproducing limitations from its source dataset. It is crucial to evaluate task accuracy, safety behavior, refusal behavior, and style strength thoroughly before deployment.