| # stage_specialized_recurrence1 | |
| ## Research artifact notice | |
| This upload is a research artifact from `recurrent-staged-loras`; validate behavior before any downstream usage. | |
| ## Run metadata | |
| - Base model: `Qwen/Qwen3-8B` | |
| - Baseline family: `stage_specialized_recurrence` | |
| - Recurrence mode: `stage_specialized` | |
| - Adapter settings: `{"latent_refiner": {"adapter_sharing": "per_step", "enabled": true, "hidden_size": 0, "num_steps": 3, "recurrence_mode": "stage_specialized"}, "latent_refiner_adapter": {"alpha": 16, "dropout": 0.0, "enabled": true, "rank": 8, "target_modules": ["refiner_proj"]}, "standard_lora": {"alpha": 32, "dropout": 0.05, "enabled": false, "rank": 16, "target_modules": ["q_proj", "k_proj", "v_proj", "o_proj", "up_proj", "down_proj", "gate_proj"]}}` | |
| - Dataset: `metamath_qa` split `train` | |
| - Training seed: `11` | |
| ## Loading | |
| Primary weights are exported as Hugging Face-compatible safetensors (single-file or sharded with index). | |
| PyTorch checkpoint artifacts (`checkpoint.pt`) are removed after safetensors export+validation. |