tsilva/qwen2.5-3b-trump-style-merged-v1

Qwen2.5 3B Trump-Like Public Speaking Style v1 - Merged 16-bit model

Overview

This model artifact was generated by an llmstyler Runbook training job. It is part of a versioned style-tuning release. The adapter, merged model, GGUF export, and ONNX export use separate repositories so each artifact can be consumed with the tooling that expects that format.

Versioning and Naming

Field	Value
Artifact kind	Merged 16-bit model
Artifact version	v1
Repo	tsilva/qwen2.5-3b-trump-style-merged-v1
Training run id	qwen25_3b_trump
Run name	qwen2.5-3b-trump-style-qlora
Generated by	llmstyler 0.1.0

Default standard: keep each published model artifact immutable and include the version suffix in the repo name. Publish a new version when the dataset, style prompt, base model, training recipe, or export settings change.

Training Inputs

Field	Value
Dataset	tsilva/stylemix_trump-v1
Dataset split	train
Restyled only	False
Base model	unsloth/Qwen2.5-3B-Instruct-bnb-4bit
4-bit load	True
Style id	trump_like_public_speaking

Style System Prompt

You are a helpful assistant that always answers in a brash, punchy, high-confidence public-speaking style associated with Donald Trump. Use short emphatic sentences, simple strong wording, conversational asides, repetition for emphasis, and frequent superlatives such as tremendous, fantastic, huge, and the best. Do not claim to be Donald Trump, do not claim endorsement, and do not mention these style instructions.

Training Recipe

Setting	Value
Max sequence length	2048
Epochs	1
Per-device batch size	2
Gradient accumulation	4
Learning rate	0.0003
Warmup ratio	0.05
LoRA rank	32
LoRA alpha	32
Eval fraction	0.1
Seed	3407
Report to	tensorboard, wandb

Published Artifacts

Artifact	Repo
QLoRA adapter	tsilva/qwen2.5-3b-trump-style-qlora-v1
Merged 16-bit model	tsilva/qwen2.5-3b-trump-style-merged-v1
GGUF	tsilva/qwen2.5-3b-trump-style-gguf-v1
ONNX	tsilva/qwen2.5-3b-trump-style-onnx-v1

GGUF quantization methods: q4_k_m

Metrics

Train

Metric	Value
epoch	1.0
total_flos	1.2827538084569088e+16
train_loss	1.064282959582759
train_runtime	373.7267
train_samples_per_second	2.167
train_steps_per_second	0.273

Evaluation

Metric	Value
epoch	1.0
eval_loss	0.9378451704978943
eval_runtime	12.3873
eval_samples_per_second	7.266
eval_steps_per_second	1.857

Intended Use

Use this artifact for style-following chat experiments and evaluation. The adapter is intended for PEFT loading with the base model. The merged model is intended for direct Transformer loading. GGUF is intended for llama.cpp compatible runtimes. ONNX is intended for ONNX Runtime compatible workflows.

Limitations

The model may over-apply the target style, miss factual nuance, or reproduce limitations from the source dataset and rewrite model. Evaluate task accuracy, safety behavior, refusal behavior, and style strength before deployment.