kddubey

cappr. Completion After Prompt Probability. Make your LLM make a choice

mpnrl. Multiple Positives and Negatives Ranking Loss

pretrain-on-test. Finetuning Mistral 7B 2,000 times + BERT and GPT-2 135,000 times, for science. Appeared at GenBench Workshop, EMNLP 2024

backprompt. Lazy KV cache data structure for language models

blog. A list of my open source contributions, and ML Qs and As