code2lora/code2lora-gru

1---2license: mit3tags: [code, lora, hypernetwork, peft, recurrent]4---5 6# Code2LoRA-GRU — streaming hypernetwork7 8Final checkpoint of the **streaming Code2LoRA-GRU** used in the paper. A91-layer GRU rolls the recurrence over per-commit diff embeddings and emits10a rank-16 LoRA adapter for `Qwen/Qwen2.5-Coder-1.5B` at *O(1)* per commit.11 12## Files13 14| File | Description |15|---|---|16| `code2lora_gru.pt`  | Trained GRU + `Code2LoRAHead` weights (~2.85 GB, fp32). |17| `metrics.jsonl`     | Per-step training metrics (loss, val EM/EditSim/CodeBLEU). |18 19## Training recipe20 21* 3 epochs of truncated BPTT (window K=16) on22  `code2lora/code2lora-data-smartcap` (train QnAs) plus23  `code2lora/code2lora-data-commits` (commit metadata + diff embeddings).24* AdamW + cosine schedule, max-seq-len 8192, bf16, single H100 80 GB.25 26## Companion model27 28`code2lora/code2lora-direct` -- the static-snapshot variant.29