| { | |
| "starting_checkpoint": "../checkpoints/pretrained/ckpt.pt", | |
| "epochs": 3, | |
| "effective_batch_size": 32, | |
| "learning_rate": 6.25e-05, | |
| "warmup_fraction": 0.002, | |
| "schedule": "linear warmup then linear decay, matching GPT-1 task adaptation", | |
| "lm_coefficient": 0.5, | |
| "classifier_dropout": 0.1, | |
| "weight_decay": 0.01, | |
| "betas": [ | |
| 0.9, | |
| 0.999 | |
| ], | |
| "epsilon": 1e-08, | |
| "grad_clip": 1.0, | |
| "max_length": 512, | |
| "seed": 42, | |
| "cpu_threads": 2, | |
| "desktop_reserve_gib": 2.5, | |
| "max_process_gib": 9.0, | |
| "checkpoint_interval": 500, | |
| "log_interval": 25, | |
| "max_consecutive_nonfinite": 8, | |
| "tasks": { | |
| "sst2": { | |
| "microbatch_size": 16, | |
| "gradient_checkpointing": false, | |
| "eval_interval": 1000 | |
| }, | |
| "race_middle": { | |
| "microbatch_size": 2, | |
| "gradient_checkpointing": true, | |
| "eval_interval": 400 | |
| } | |
| } | |
| } | |