Personal-style LoRA

The goal is narrow and hard: make a model that writes the way I write, and be able to prove it did not just memorize me.

I fine-tuned and RL-post-trained Qwen3-8B and Qwen3-14B LoRA models — SFT, KTO, DPO, and on-policy multi-turn GRPO — on a 7.0M-word corpus of my own writing, with LoRA rank 64 and alpha 128. A pinned Qwen3-32B teacher generates synthetic training dialogue so the instruction format does not drift away from the underlying style.

The evaluation work had to come before the training work. idbench-v1 is an identity benchmark of 1,448 graded responses across 12 categories, scored six ways automatically — style divergence, authorship rank, memorization, privacy leakage among them — and then reviewed blind by humans. Anything that wins the automatic metrics but reads as imitation in blind review does not ship.

Two results shaped the rest. First, a rejection-sampled fact curriculum raised the idbench macro score by +0.08 with no regressions: the model got more of my actual positions and habits right, not just my punctuation. Second, on-policy DPO repair with 3,800 prompts × 8 samples generated across 7 GPUs in about 25 minutes, scored and trained as a rank-16 LoRA, gated by blind A/B review — that is the loop that actually moves a stubborn failure mode.

Two supporting systems matter as much as the training. I fit and streamed a Jacobian lens over the merged LoRA checkpoint to read out what the model does internally, behind a parity gate against the reference implementation. And privacy is fail-closed: a sha-pinned GGUF entity detector behind a hand-written C++17 persistent bridge scrubs personal data before anything reaches training.

What I'd do differently: define what a good response looks like to a human reader before building more automated metrics. Metrics that are easy to compute have a way of becoming the objective, and on a task this personal the ground truth is a reader, not a score.