tsilva/qwen2.5-3b-trump-style-merged-v2

Qwen2.5 3B Trump-Like Public Speaking Style v2 - 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	v2
Repo	tsilva/qwen2.5-3b-trump-style-merged-v2
Training run id	qwen25_3b_trump_v2
Run name	qwen2.5-3b-trump-style-qlora-v2
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-v2
Dataset split	train
Restyled only	True
Base model	unsloth/Qwen2.5-3B-Instruct-bnb-4bit
4-bit load	True
Style id	trump_like_public_speaking

Style System Prompt

Training Recipe

Setting	Value
Max sequence length	2048
Epochs	5
Per-device batch size	2
Gradient accumulation	4
Learning rate	0.0003
Warmup ratio	0.05
LoRA rank	32
LoRA alpha	32
Eval fraction	0.2
Seed	3407
Report to	tensorboard, wandb

Published Artifacts

Artifact	Repo
QLoRA adapter	tsilva/qwen2.5-3b-trump-style-qlora-v2
Merged 16-bit model	tsilva/qwen2.5-3b-trump-style-merged-v2
GGUF	tsilva/qwen2.5-3b-trump-style-gguf-v2
ONNX	tsilva/qwen2.5-3b-trump-style-onnx-v2

GGUF quantization methods: q4_k_m

Metrics

Train

Metric	Value
epoch	2.4444444444444446
total_flos	4889840584826880.0
train_loss	1.3235322819514708
train_runtime	251.3496
train_samples_per_second	2.865
train_steps_per_second	0.358

Evaluation

Metric	Value
epoch	2.4444444444444446
eval_loss	1.3690619468688965
eval_runtime	4.1924
eval_samples_per_second	8.587
eval_steps_per_second	2.147

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.