#model training

2 articles tagged with "model training"

How DPO Cuts Text Degeneration by 59% Without Retraining From Scratch

DharmaOCR's methodology proves Direct Preference Optimization isn't just for chat alignment. Applied after supervised fine-tuning, DPO reduced text degeneration by an average of 59.4% across five vision-language model families—with zero exceptions.

Dr. Sana Okafor Jun 3, 2026

research 9 min read

Text Degeneration in LLMs: The Hidden Production Cost Inflating Inference by 42%

A structural failure mode in autoregressive language models causes fewer than 3% of requests to consume nearly half of total inference time. New research from DharmaOCR shows the problem is built into training objectives—and proposes a fix grounded in the training distribution itself.

Dr. Sana Okafor May 22, 2026