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.
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)
