EP.04 / PROJECT CASE FILE / 01
Chatbot Arena Preference Prediction
Fine-tuned and ensembled open-source LLMs for human preference prediction, placing 67th out of 1,849 teams.
A closer look at the problem, engineering decisions, system design, and measurable outcome.
01 / Context
Predict which of two chatbot responses a human evaluator would prefer in the LMSYS Chatbot Arena competition.
02 / Approach
Fine-tuned Gemma-2-9B and Llama-3.1-8B with 4-bit QLoRA, inserting LoRA adapters into attention layers while working within limited GPU memory.
03 / Engineering
Tuned learning rate, prompt length, and frozen layers for stable training, then sorted test samples by length to reduce padding during multi-GPU inference.
04 / Model Strategy
Combined complementary model outputs in a weighted ensemble to improve generalization and preference-prediction accuracy.
05 / Result
Earned a Kaggle Silver Medal, finishing 67th of 1,849 teams—top 4% overall.