We trained a personal voice DoRA on Qwen3-8B for $1.50

Aiconic trained a personal voice DoRA adapter on Qwen3-8B using 6,128 Telegram message pairs extracted from a single user’s chats, spending $1.50 on a Vast.ai RTX 3090 instance [devto]. The model beat the base Qwen3-8B in every round of blind A/B testing, with raters consistently preferring its tone and phrasing. Dataset curation capped 12 message pairs per chat to prevent bias from high-volume threads, leaving 6,128 training and 322 validation pairs [aiconic].

The adapter used Weight-Decomposed Low-Rank Adaptation (DoRA), which splits pretrained weights into magnitude and direction, applying LoRA-style updates only to direction while training magnitude separately. Training ran for 3 epochs with a batch size of 2, gradient accumulation of 8, and a learning rate of 2e-4 using a cosine scheduler with 50 warmup steps. The setup targeted q_proj, k_proj, v_proj, and o_proj modules via Hugging Face Transformers and Peft, with rank 16 and Lora alpha 32.

DoRA showed no catastrophic forgetting across 50 general-knowledge tasks—unlike standard fine-tuning, which often erodes base capabilities. The method preserved Qwen3-8B’s core performance while injecting personal voice. At $1.50 and under 3 hours, this approach makes individual-level voice cloning viable outside labs. Most consumer AI still uses segment-level personalization (e.g., age, region), but this shows individual tuning can be cheap, fast, and effective.

The real win isn’t just cost—it’s control. Users keep their data private, train locally, and avoid re-uploading models. That flips the script from Big Tech’s centralized personalization, where user data fuels opaque models.