A Next-Gen Code Model

Code Intelligence.
Simplified.

Curated code datasets, QLoRA fine-tuning, and robust evaluations — all in one open-source package.

Enter the flow →✦ Try the demo
Training Run #1024
In Progress...

Explore the Project

Each section dives deep into a different aspect of CodeSage.

Interactive

Live Playground

Code completion, review, test generation, and documentation — powered by Gemini.

Explore
Pipeline

Training Deep-Dive

Dataset curation, QLoRA config, W&B tracking, adapter merging, Modal deployment.

Explore
Results

Benchmarks

HumanEval, MBPP, HumanEval+, SQL generation — animated charts and comparisons.

Explore
Architecture

Inference Stack

vLLM + PagedAttention: 12,500 tok/s, interactive architecture diagram.

Explore
Analysis

What Breaks

6 real failure modes — catastrophic forgetting, OOM, quantization noise, and more.

Explore
Integration

MCP Server

4 tools exposed via Model Context Protocol — usable by Claude, Cursor, any agent.

Explore

Core Tech Stack

PythonPython
TypeScriptTypeScript
ReactReact
Next.jsNext.js
TailwindTailwind
PyTorchPyTorch
vLLMvLLM

Resume BulletTL;DR

Fine-tuned Llama 3.3 8B on 52K curated code instruction pairs using QLoRA (r=16, DoRA, all-linear targets, NF4 4-bit quantization) with Unsloth on Modal A100 80GB for $9.40 total; achieved +7.1pp HumanEval pass@1 and +6.3pp MBPP pass@1 over base model through dataset curation (3-pass syntax/execution/deduplication filtering), mixed training objective preventing catastrophic forgetting, and post-merge evaluation methodology; deployed with vLLM (PagedAttention, continuous batching, 12,500 tok/s on H100, 24x vs naive serving) on Modal serverless GPU; exposed via MCP server with 4 tools (complete, review, test-generate, explain) usable by Claude Desktop, Cursor, and any LangGraph agent.