Media Bias Detection

Political slant in news is often carried by framing rather than by explicit claims, which makes it a useful test of how well language models understand text.

Fine-tuning

LLaMA-3 8B and Gemma-2B were fine-tuned with LoRA on a labelled news dataset, reaching 96% accuracy across five political-bias categories.

Cost versus quality

Benchmarking compute–performance trade-offs identified a configuration that cut inference cost by 34% at near-constant accuracy.