gemma-3

A clean, minimal, and educational implementation of the Gemma 3 language model architecture using pure PyTorch and JAX

AI-powered coding assistant that sits on top of your IDE. Screenshot any coding problem, get instant AI solutions via Google Gemini.

Experiments in Latin dactylic hexameter generation with transformers: A hybrid post hoc feedback framework

Private, 100% on-device personal AI assistant for Android powered by Gemma 3 1B & LiteRT-LM. 51 native system tools, offline SQLite memory, and zero cloud telemetry.

High-performance RLHF/GRPO pipeline scaling Gemma 3 on GKE Ray Clusters (B200/H200) using NVIDIA NeMo-RL. Includes native FSDP checkpoint merging and zero-shot vLLM benchmarking.

Do language models understand the physics of molten steel? SLPB-50 benchmarks physical plausibility and constraint blindness in pyrometallurgical AI.

OpenRouter free-LLM CLI & skill for coding agents

A local-first, AI-powered knowledge architecture system that transforms raw study materials into professional-grade study guides using Gemma 3 on Apple Silicon.

[EE6406]ANALYTIC & ENSEMBLE LEARNING Project:End-to-end toolkit for medical SFT dataset preparation, Gemma 3 4B IT fine-tuning (LoRA/DoRA/QLoRA), vLLM inference, and HealthBench evaluation.

Full fine-tune google/gemma-3-1b-it updated for Bangla news summarization on XL-Sum Bengali. chrF++ 24.91 against 22.78 zero-shot, ROUGE-L 11.43 against lead-3’s 10.34. Reported against both baselines, because a summarizer that cannot beat lead-3 is not summarizing.

Measures whether an LLM's self-reports agree with what its internals show. Injected concepts are perfectly recoverable by probes (AUC 1.000) and absent from the model's verbal report - report-state agreement fails in a graded way. Gemma-3-4B, Jacobian lens, ~3,750 trials.

QLoRA adapter finetuned from gemma-3-4b-it for retrieval-grounded, bilingual EN/ID international-trade assistance. Runs offline in place of a hosted RAG API.

Fish Audio Drama 3 int4, Korean/English expert-pruned (256→160) · runs within 24 GB VRAM · weights on Hugging Face

Offline, audio-first nature scavenger hunt agent powered by local multimodal Gemma 3 vision.

Production-ready RAG talent intelligence engine with Google Gemma 3, Supabase PostgreSQL, in-memory parsing, and dual-mode conversational search.

AI-powered DSA tutor using Gemma 3 — step-by-step algorithm traces, Big-O derivations via recurrence relations, multi-language code generation, and adaptive practice problems.

ResearchNexus AI: A Gemma 3-powered academic workstation. Built with 95% Vibe Coding.

A production-ready Multimodal Misinformation Analyzer API. Built with FastAPI and Docker, it leverages a fine-tuned Gemma-3 model via LoRA adapters to detect fake news by cross-referencing textual claims with visual context

On-device OSHA construction safety violation detector — Gemma 4 E4B + Unsloth + Ollama. Gemma 4 Good Hackathon submission (Global Resilience track).

Turn any PDF, scan, or photo into clean structured Markdown with Google's Gemma 3 27B vision model on Nebius Token Factory. Streaming Streamlit OCR app with real table extraction.