Draft Voice

When I am doing much of my writing I dislike how LLM's writing assistance tends towards that models style rather than my own. I often like to have agents throw together a draft of something but it sounds so unlike me that its less inspiration and more something I have to rewrite from scratch.

To help me write better drafts and to play around with what I can do with limited training data I have been training a LoRA adapter on my own text. I have scraped all my writing into a dataset.

I have been augmenting it in several ways. I really would like to have steerable responses so I have added a prompt to every training item. So far for adapter v2 it is not that steerable. It tends to rewrite in my voice only on short sentences. It keeps the length of the input that its supposed to rewrite instead of the target length.

To test the adapters I give Opus 5.5 some examples of my writing. I then present it with the adapter and my real writing to test if it cannot tell the difference. If Opus picks the adapter vs me 50/50 then I am confident that the adapter sounds like me. As you can see from this data though v2 did not improve much over v1 in prose even though v2 improved slightly on steerablity.

Length of your real text	Run 2 spotted	Run 1 spotted
Under 50 words	43%: can't tell	53%
50–150 words	74%	76%
150–400 words	89%	91%
Over 400 words	94%	96%
Overall	75%	79%