https://rajathpatel23.github.io/tags/sft/Recent content in SFT on Rajat PatelHugo -- 0.167.0en-usrpatel12@umbc.edu (Rajat Patel)rpatel12@umbc.edu (Rajat Patel)Sun, 27 Sep 2026 00:00:00 +0000https://rajathpatel23.github.io/posts/tiny-shakespeare-sft-dpo/Sun, 27 Sep 2026 00:00:00 +0000rpatel12@umbc.edu (Rajat Patel)https://rajathpatel23.github.io/posts/tiny-shakespeare-sft-dpo/A 15M-parameter model pretrained on tiny-shakespeare, fine-tuned to stop after one line, then aligned with DPO. The same learning rate that worked for SFT broke generation. A tenfold cut fixed it.