| { | |
| "base_model": "HuggingFaceTB/SmolLM2-135M-Instruct", | |
| "dataset_name": "rogue-security/prompt-injections-benchmark", | |
| "output_dir": "./foresight-135m", | |
| "test_split_ratio": 0.1, | |
| "seed": 42, | |
| "num_epochs": 5, | |
| "per_device_train_batch_size": 4, | |
| "per_device_eval_batch_size": 8, | |
| "gradient_accumulation_steps": 8, | |
| "learning_rate": 2e-05, | |
| "weight_decay": 0.01, | |
| "warmup_ratio": 0.05, | |
| "lr_scheduler_type": "cosine", | |
| "max_grad_norm": 1.0, | |
| "max_seq_len": 512, | |
| "logging_steps": 25, | |
| "eval_steps": 100, | |
| "save_steps": 200, | |
| "save_total_limit": 2, | |
| "load_best_model_at_end": true, | |
| "metric_for_best_model": "eval_loss", | |
| "greater_is_better": false | |
| } |