train_config.json
715 B · 29 lines · json Raw
1 {
2 "model": "teutonic-sft-full-v2-e1 (final)",
3 "mix": "data/mixes/topup-longctx-v1",
4 "output_dir": "outputs/teutonic-sft-topup-longctx-v1",
5 "max_length": 65536,
6 "kernels": "hub",
7 "attn_implementation": "auto",
8 "per_device_train_batch_size": 1,
9 "gradient_accumulation_steps": 4,
10 "learning_rate": 1e-05,
11 "lr_scheduler_type": "cosine",
12 "warmup_ratio": 0.05,
13 "min_lr_rate": 0.1,
14 "weight_decay": 0.0,
15 "max_grad_norm": 1.0,
16 "num_train_epochs": 1.0,
17 "save_steps": 40,
18 "save_total_limit": 4,
19 "snapshot_fractions": [],
20 "eval_steps": 20,
21 "logging_steps": 1,
22 "seed": 20261011,
23 "dataset_num_proc": 16,
24 "report_to": [
25 "wandb"
26 ],
27 "run_name": "teutonic-sft-topup-longctx-v1"
28 }
29