| # Sweep Experiment Scripts (Jul 9 2026) | |
| All LR sweep scripts for FFT on `generated_data.jsonl` (1314 real API texts). | |
| ## File Naming Convention | |
| - `train_per_token_vali_overfi_roberta_lr*.py` — RoBERTa-large fresh HF backbone, 4 LR sweep | |
| - `train_per_token_vali_overfi_deberta_lr*.py` — DeBERTa-v3-large SN32 weights, 5 LR sweep | |
| - `train_per_token_vali_trndffi_overfi_lr*.py` — RoBERTa-large transfer (from synthetic sandwich ckpt), 5 LR sweep | |
| - `train_per_token_vali_overfi_base.py` — Clean base template (patience=1000, epoch ckpts) | |
| - `train_per_token_vali_overfi_deberta.py` — DeBERTa SN32 base template | |
| - `train_per_token_vali_trndffi_overfi.py` — Transfer base template | |
| ## Results Summary | |
| | Config | LR | Best Reward | | |
| |--------|----|-------------| | |
| | RoBERTa fresh | 1e-5 | 0.8347 | | |
| | RoBERTa trndffi | 1e-5 | 0.8570 ← winner | | |
| | RoBERTa trndffi | 1e-6 | 0.8005 | | |
| | DeBERTa SN32 | any | ~0.37 (F1=0 all LRs) | | |
| See `AGENTS_BUCKT_TRAIN.md` section 16 for full analysis. | |