| --- |
| license: apache-2.0 |
| library_name: transformers |
| tags: |
| - code |
| - code-verification |
| - qwen3_5 |
| - ms-swift |
| --- |
| |
|  |
|
|
| **Code verification from execution and repair evidence** |
|
|
| [Why CodeVerifier](#-why-codeverifier) · [Highlights](#-highlights) · [Models](#-models) |
|
|
|  |
|  |
|
|
| > [!NOTE] |
| > Three size variants are provided in `2B/`, `4B/`, and `9B/`. |
|
|
| ## 🌐 Why CodeVerifier? |
|
|
| Execution provides reliable feedback, but it often arrives too late for large candidate pools, online policy updates, and repository workflows. CodeVerifier supplies timely reward signals for reinforcement learning with verifiable rewards (RLVR) and candidate search, as well as rapid quality feedback for filtering generated training data before a full test run. |
|
|
| Rather than returning only a scalar score, CodeVerifier produces a verdict, a brief explanation, and supporting code regions for complete programs, intermediate edits, and repository code states. It learns these judgments from historical execution outcomes and successful repairs through evidence projection and RVPG. |
|
|
| ## ✨ Highlights |
|
|
| | | | |
| |---|---| |
| | **Structured judgments** | Verdict, explanation, and line-level evidence in one response. | |
| | **Evidence-projected supervision** | Execution outcomes and repair evidence become structured training targets. | |
| | **Region-aware optimization** | Verdict and evidence feedback are routed to their corresponding output fields. | |
|
|
| ## 🧭 How it works |
|
|
| 1. Pair each code state with its execution outcome and repair-associated regions. |
| 2. Project those signals into structured targets for supervised fine-tuning. |
| 3. Refine the verifier with region-aware online optimization. |
|
|
| ## 📦 Models |
|
|
| | Model | Folder | |
| |---|---| |
| | CodeVerifier-2B | `2B/` | |
| | CodeVerifier-4B | `4B/` | |
| | CodeVerifier-9B | `9B/` | |
|
|
