README.md

---
license: apache-2.0
base_model: unsloth/Qwen3.6-35B-A3B
tags:
- forge-optimizer
- backend-code-optimization
- insforge
- grpo
- unsloth
---

# forge-optimizer

A LoRA fine-tune of **Qwen3.6-35B-A3B** (MoE) that turns unoptimized backend code into
efficient, correct code — the skill measured by [forger-bench](https://github.com/Hcoder10/forger-bench),
an efficiency-aware benchmark for AI-generated InsForge SDK code.

Trained with **Unsloth** (bf16 LoRA) + an agentic GRPO loop adopting **CUDA-Agent**
(arXiv 2602.24286): the model writes a solution, the forger-bench grader runs+verifies+
measures real server metrics, and a discrete milestone reward (-1 incorrect/scaleBug, 1
wasteful, 2 beats-naive, 3 near-optimal) drives RL.

## Contamination control
Never trained on a sealed test task; held-out concepts measure optimization skill vs
template memorization.