| license: mit | |
| tags: [code, lora, hypernetwork, peft] | |
| # Code2LoRA — direct-projection hypernetwork | |
| Final checkpoint of the **direct-projection** Code2LoRA hypernetwork used in | |
| the paper. Maps a repository-level embedding into a rank-16 LoRA adapter for | |
| `Qwen/Qwen2.5-Coder-1.5B` in a single forward pass. | |
| ## Files | |
| | File | Description | | |
| |---|---| | |
| | `code2lora_direct.pt` | Trained `Code2LoRAHead` weights (~2.7 GB, fp32). Loaded with `torch.load(map_location="cpu")`. | | |
| ## Training recipe | |
| * 3 epochs on the `code2lora/code2lora-data-snapshots` dataset. | |
| * AdamW + cosine schedule, max-seq-len 8192, bf16, single H100 80 GB. | |
| * See [`code2lora/code2lora`](https://github.com/) for the trainer code. | |
| ## Companion model | |
| `code2lora/code2lora-gru` -- the streaming-recurrent variant trained on | |
| commit deltas. | |