Legesher

Code in your native language.

Legesher is an open-core toolchain that lets developers create, innovate, and collaborate on code in their own language. Existing tooling, CI, and code review keep working regardless of the language you use, so code in the language of your ideas.

• 🌐 Website: legesher.io
• 💻 GitHub: github.com/Legesher
• 💬 Slack: legesher.io/go/slack

---

What lives here on Hugging Face

This org hosts the research artifacts behind Legesher. Featuring the Language Decoded project, we highlight the impact of native code in the performance of multilingual models as our contribution to Cohere's Tiny Aya Expedition.

📊 Datasets

• language-decoded-data — multilingual code corpora across experimental conditions (en / zh / es / ur)
• language-decoded-community — natively-authored code contributed by developers
• language-decoded-experiments — training logs, configs, and evaluation results

🤖 Models

LoRA adapters fine-tuned on CohereLabs/tiny-aya-base across conditions:

• Condition 1 (english code baseline)
• Condition 2 (translated code identifiers via Legesher)
• Condition 3 (native code)
• Condition 5 (fully translated code via Legesher + Aya Expanse)

🎮 Spaces

• legesher-native-code — submit natively-authored code to the community corpus

---

The product side

Legesher ships 51 experimental Python language packs (am, ar, bn, … zh) plus a VS Code extension and full LSP server. Translation happens through git filters, so the repo on disk stays English and your editor renders your native language. See the main repo for installation and the language pack list.

---

Built by Madison Edgar and the Legesher community. Apache-2.0 for code; see individual repos for dataset and model licenses.