Where it started
A companion model may invent a user's personal experiences. Reducing fabrication must not come at the cost of extracting information explicitly provided.
My contribution
Using Qwen3-0.6B, I prepared a small training and held-out set, trained LoRA with PyTorch, Transformers and PEFT, saved and reloaded the adapter, and reviewed before-and-after responses individually.
Finishing training is a start. Keeping the failures tells you what to do next.
Three engineering decisions
Keep the base, update the adapter
Freeze the base model and inject LoRA into attention projections. Save the adapter separately and verify reloading and unchanged base files.
Pair unknown and known questions
Check refusal on missing personal information, then test extraction when that information is supplied, rather than teaching blanket refusal.
Look beyond loss
Review answers one by one. Preserve the incorrect refusal on a note-recipient question as a regression instead of equating completed training with product readiness.
What came out of it
- A small run used 32 training examples, 8 held-out questions and 48 training steps.
- 573,440 adapter parameters were trained, saved and reloaded in a fresh process.
- Some unknown-information answers fabricated less, but a known-information question regressed to an incorrect refusal. Product quality remains unaccepted.
Training, saving and reloading completed; product quality not accepted
What's next
Improve training examples and evaluation coverage while balancing refusal with extraction of known information.
Project notes are grounded in reviewed source, experiment records and personal confirmations.
Content reviewed: 2026.10.28