README.md

metadata

title: FinCompress
emoji: 🗜️
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 6.6.0
app_file: app.py
pinned: false
short_description: FinBERT compression via KD, INT8 quant, and pruning

🗜️ FinCompress

Compressing FinBERT (109M params) into a 19M-parameter student using knowledge distillation, INT8 quantization, and structured attention-head pruning — all benchmarked on financial sentiment classification.

What this Space shows

• Live demo: run both the 109M teacher and 19M student side-by-side on any financial sentence
• Benchmark table: all 7 model variants (teacher, KD students, PTQ, QAT, pruned)
• Architecture explainer: how KD, quantization, and pruning each work

Links

• 📦 GitHub — FinCompress
• 🤗 Student Model Weights
• 📊 Dataset: takala/financial_phrasebank