We trained a personal voice DoRA on Qwen3-8B for $1.50 — beat stock model 100% in blind A/B
A personal voice DoRA was trained on Qwen3-8B using 6128 Telegram messages for a cost of $1.50. In blind A/B testing, the DoRA model outperformed the stock model 100% of the time, with no catastrophic forgetting observed across various tasks. Notably, in one instance, the DoRA model was rated as sounding more like the user than their actual response.
- ▪The DoRA adapter was trained on a dataset of 6128 personal Telegram messages.
- ▪The training process took approximately 3.5 hours on a single RTX 3090 GPU, costing around $1.50.
- ▪In a blind evaluation, the DoRA model was preferred over the stock model in all tests conducted.
DEV.to (Top) files mainly under programming. We currently carry 4,877 of its stories.
Opening excerpt (first ~120 words) tap to expand
try { if(localStorage) { let currentUser = localStorage.getItem('current_user'); if (currentUser) { currentUser = JSON.parse(currentUser); if (currentUser.id === 3914709) { document.getElementById('article-show-container').classList.add('current-user-is-article-author'); } } } } catch (e) { console.error(e); } Yuka Kust Posted on May 25 We trained a personal voice DoRA on Qwen3-8B for $1.50 — beat stock model 100% in blind A/B #ai #llm #machinelearning #showdev TL;DR. Trained a DoRA adapter on Qwen3-8B using 6128 personal Telegram messages. Cost: $1.50 on a single Vast.ai RTX 3090. In blind head-to-head A/B, the DoRA-tuned model beat stock Qwen3-8B 100% of the time. Zero catastrophic forgetting on 50 general-knowledge tasks.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at DEV.to (Top).
Discussion
0 commentsMore from DEV.to (Top)
