Research

PAUL Open is an applied AI research initiative that fine-tunes Google's open-weight Gemma 4 model family for two underserved intersections: multilingual Indian-language STEM education and life-sciences reasoning.

Using a disciplined two-stage pipeline (Supervised Fine-Tuning followed by Direct Preference Optimization) on a small, hand-curated, contamination-audited corpus of 245 examples across 10 Indian languages, the project investigates whether targeted behavioral specialization—Socratic tutoring, concise numerical problem-solving, clean Indic-language scientific translation, and natural pedagogical tone—can be reliably injected into a compact open model without destroying its baseline capabilities, and whether this can be done reproducibly on a single consumer GPU.

Stage 1: Supervised Fine-Tuning (SFT)

The first stage uses QLoRA to inject structured domain knowledge and basic formatting compliance using 180 highly curated examples. This focuses on preservation of baseline capabilities while aligning the model to the rigorous formatting required by the domains.

Stage 2: Direct Preference Optimization (DPO)

The second stage targets behavioral nuance using 65 preference pairs. Unlike standard DPO which often inadvertently trains for length (verbosity bias), PAUL Open's DPO v2 corrective dataset explicitly penalizes inappropriate brevity, robotic coldness, and lack of structure where verbosity and detail are necessary, while simultaneously penalizing unhelpful rambling, irrelevant tangents, and unnecessary complexity where brevity and simplicity are requested.