Fine-tuning Qwen on My Emails
I trained Qwen3.5-4B on my own cold emails with LoRA, improved the training examples, and compared the fine-tune with untuned Qwen on 30 fresh briefs.
About
I fine-tuned Qwen3.5-4B on my own cold emails to see whether a small model could learn how I write outreach. Using LoRA on Tinker, I built training examples from emails I’d sent, tested successive adapters, and revised the data when the model confused my experience with other people’s work. I then asked Astra to improve the fine-tuning process. Rewriting the same 15 examples to preserve my exact question and sign-off produced the strongest tune. On 30 fresh briefs, an LLM judge ranked it first in 25 comparisons with the earlier tune and untuned Qwen; it preserved my exact question in all 30 drafts. Five drafts still contained unsupported claims or name errors. My main takeaway was that the quality of each training example mattered more than simply adding examples, and that promising drafts needed to hold up on new recipients. I published the weights on Hugging Face and ran the model locally on my Mac.
- stack
- Qwen3.5-4BLoRATinkerHugging FaceLocal inferenceCodex
- links
- Case study ↗