README.md

---
license: mit
library_name: code2lora
tags:
  - code2lora
  - lora
  - hypernetwork
  - qwen2.5-coder
base_model: Qwen/Qwen2.5-Coder-1.5B
---

# Code2LoRA — evo multitask v2 (multitask-2)

GRU + multi-task hypernetwork head for **Code2LoRA**. Given a repository's
commit history (encoded with `Qwen/Qwen3-Embedding-0.6B`), it generates a LoRA
adapter for the frozen base model `Qwen/Qwen2.5-Coder-1.5B`.

This **v2** checkpoint extends the multi-task head so the `qa` task covers both:

* **diff QA** — questions about what a commit changed, and
* **repo-state QA** — questions about the current code (what a method/class does,
  a constant's value, a function signature, control flow, …),

in addition to the **assert_rhs** assertion-completion task.

`num_tasks = 2` (task-conditioned): `assert_rhs` (index 0), `qa` (index 1). The
same walk produces genuinely different adapters per task.

Load it in the terminal tool:

```bash
pip install code2lora
c2lt                       # uses this checkpoint by default
```

Or point any C2L deployment at it:

```bash
export C2L_CKPT_REPO=code2lora/c2l-evo-multitask-v2
export C2L_CKPT_FILE=gru_head.best.pt
```

File: `gru_head.best.pt`.