| --- |
| license: mit |
| base_model: Qwen/Qwen2.5-Coder-1.5B |
| tags: |
| - code2lora |
| - hypernetwork |
| - lora |
| - code |
| library_name: code2lora |
| --- |
| |
| # code2lora_cli_multitask |
|
|
| Multi-task **Code2LoRA-GRU** hypernetwork checkpoint used by the `c2l-terminal` |
| (`code2lora` pip package). Given a repository's commit history it generates a |
| LoRA adapter for a frozen **Qwen2.5-Coder-1.5B** base, for two auto-routed tasks: |
|
|
| - `assert_rhs` (task 0) — test-assertion completion |
| - `qa` (task 1) — free-form repository question answering (diff + repository-state) |
|
|
| ## Training |
|
|
| - Warm-started from the single-task assertion GRU hypernetwork (grown 0→2 tasks |
| via a zero-padded, output-identical initialization). |
| - Finetuned on **164,173** commit-grounded Q&A pairs (73,829 diff-grounded + |
| 90,344 repository-state) **+ a 10% assertion-completion rehearsal sample**, |
| which prevents catastrophic forgetting of the original task. |
|
|
| ## Evaluation (held-out test splits, CR / IR) |
|
|
| | task | metric | CR | IR | |
| |---|---|---|---| |
| | assertion completion | exact match | 60.5 | 66.1 | |
| | repository QA | QA quality (½·(ROUGE-L+token-F1)) | 0.229 | 0.230 | |
|
|
| Assertion accuracy matches the original single-task checkpoint (60.3 / 64.5) |
| while adding the QA capability at no measurable cost. |
|
|
| ## Usage |
|
|
| ```bash |
| pip install code2lora |
| C2L_CKPT_REPO=code2lora/code2lora_cli_multitask c2lt |
| ``` |
|
|
