Per-Token PVI Analysis

Starting the model… the first request after an idle period takes ~20–30s while the model loads onto the GPU.

Product Title

Product Description

Examples

Methodology. For each word token in the listing, we remove it and recompute PVI on the ablated text. The difference ΔPVI = PVI(with) − PVI(without) measures how much the token contributes to the listing’s machine-usable information. Positive Δ means the token helps the AI agent’s decision; negative Δ means it hurts. Tokens are extracted via word boundaries and optionally filtered to three established NLP lexicons: Opinion Lexicon (Hu & Liu 2004), VADER (Hutto & Gilbert 2014), and Empath business/money/shopping categories (Fast et al. 2016).

Implementation. The base model (Llama-3.1-8B-Instruct) is quantized to 4-bit (NF4) and served with two LoRA adapters (weights) on a single T4 GPU via Modal. Anonymized request logs (IP address, inputs, and results) are retained for research purposes. Repeated queries are served from cache.