| license: mit | |
| tags: [code, lora, hypernetwork, peft, recurrent] | |
| # Code2LoRA-GRU — streaming hypernetwork | |
| Final checkpoint of the **streaming Code2LoRA-GRU** used in the paper. A | |
| 1-layer GRU rolls the recurrence over per-commit diff embeddings and emits | |
| a rank-16 LoRA adapter for `Qwen/Qwen2.5-Coder-1.5B` at *O(1)* per commit. | |
| ## Files | |
| | File | Description | | |
| |---|---| | |
| | `code2lora_gru.pt` | Trained GRU + `Code2LoRAHead` weights (~2.85 GB, fp32). | | |
| | `metrics.jsonl` | Per-step training metrics (loss, val EM/EditSim/CodeBLEU). | | |
| ## Training recipe | |
| * 3 epochs of truncated BPTT (window K=16) on | |
| `code2lora/code2lora-data-smartcap` (train QnAs) plus | |
| `code2lora/code2lora-data-commits` (commit metadata + diff embeddings). | |
| * AdamW + cosine schedule, max-seq-len 8192, bf16, single H100 80 GB. | |
| ## Companion model | |
| `code2lora/code2lora-direct` -- the static-snapshot variant. | |