| --- |
| title: RustGen |
| emoji: 🦀 |
| colorFrom: gray |
| colorTo: red |
| sdk: gradio |
| sdk_version: 6.18.0 |
| app_file: app.py |
| pinned: false |
| short_description: Paste Python, get execution-verified Rust |
| --- |
| |
| # RustGen |
|
|
| Teaching a small code model a new language. You give it a Python function or an |
| English description; a vanilla **Qwen2.5-Coder-1.5B** writes the Rust, and `rustc` |
| compiles and tests the result before it is shown. When the first attempt fails to |
| compile, a **compile-guided cascade** retries with a retrieved idiom exemplar plus |
| similar examples — retrieval is only used to recover compile failures, never to |
| inflate the output blindly. |
|
|
| ## What it does |
|
|
| - **Python → Rust** — the primary task; the model translates a Python function. |
| - **English → Rust** — direct, or via a Python pivot (Qwen drafts Python first). |
| - **Verification** — every output is compiled and run against tests; paste your own |
| asserts in *Reference tests* to check against ground truth. |
|
|
| ## Results (MultiPL-E humaneval-rs, 156 problems, same harness) |
|
|
| | System | pass@1 | |
| |---|---| |
| | Qwen-1.5B vanilla | 37.8% | |
| | + compile-guided cascade | 44.9% | |
| | + compiler-feedback repair | 46.2% | |
| | Qwen-7B-4bit + full pipeline | 62.8% | |
|
|
| The live demo runs the 1.5B cascade. The 7B figures are measured offline. |
|
|
| ## Notes |
|
|
| - On a free CPU Space, generation takes a while per attempt (the cascade may run a |
| few). Upgrade the Space hardware to a GPU for interactive speed. |
| - The models are downloaded from the Hub on first use; no token required. |
|
|
