Upload 06b_bf16_r16/train_args.json with huggingface_hub
Browse files- 06b_bf16_r16/train_args.json +32 -0
06b_bf16_r16/train_args.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"base": "Qwen/Qwen3-0.6B",
|
| 3 |
+
"corpus": "data/distill/train_structured.jsonl",
|
| 4 |
+
"out": "results/loop9h/p3_m1",
|
| 5 |
+
"precision": "bf16",
|
| 6 |
+
"mode": "lora",
|
| 7 |
+
"lora_r": 16,
|
| 8 |
+
"targets": "dense",
|
| 9 |
+
"max_len": 1024,
|
| 10 |
+
"batch": 8,
|
| 11 |
+
"grad_accum": 3,
|
| 12 |
+
"max_steps": 2000,
|
| 13 |
+
"lr": 0.0002,
|
| 14 |
+
"warmup": 30,
|
| 15 |
+
"min_lr_frac": 0.1,
|
| 16 |
+
"weight_decay": 0.0,
|
| 17 |
+
"permute_prob": 0.5,
|
| 18 |
+
"kd_alpha": 1.0,
|
| 19 |
+
"kd_temp": 2.0,
|
| 20 |
+
"no_grad_checkpoint": false,
|
| 21 |
+
"vram_cap_gb": 8.0,
|
| 22 |
+
"save_steps": 250,
|
| 23 |
+
"seed": 0,
|
| 24 |
+
"limit_rows": 0,
|
| 25 |
+
"deadline": 1790185384.0,
|
| 26 |
+
"trainable_params": 10092544,
|
| 27 |
+
"steps_done": 2000,
|
| 28 |
+
"steps_this_run": 2000,
|
| 29 |
+
"wall_s": 4478.817822694778,
|
| 30 |
+
"sps_steady": 2.2359451534156514,
|
| 31 |
+
"hit_deadline": false
|
| 32 |
+
}
|