Calibrated accommodation pairs reward dialect matching in post-training
Original title: Answering in the User's Variety: Calibrated Accommodation Pairs for Post-Training
Across 10 open lineages, adjusted DPO still favored British and Australian spelling over AAE and Singlish.
The paper traces how supervised fine-tuning, DPO, and reinforcement learning change whether chat models match a user's dialect. Post-training standardizes replies to informal cues, while models still produce AAE features when asked. Calibrated accommodation pairs reward matching dialect density without adding dialect for Standard-English users or hurting instruction following.
- Content-matched reply pairs separate matching and non-matching answers.
- CAP improved unseen-feature convergence on a fully open model.
- One-sided pairs gained similar matching only by overshooting the user.
- Authors release code for the CAP accommodation method.
From the original article
1 min read at huggingface.co
People adapt how they talk to the person in front of them. However, chat models tend to respond in Standard English regardless of the dialect used by the user. A uniform answer can hide whether training treats these users alike, and output audits cannot tell as well.
By Zhizhuo Kou, Heming Li, Jian Yang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.