File size: 2,113 Bytes
35d483e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
{
  "run": {
    "name": "partial-source-holdout-tinytcn",
    "output_dir": "artifacts/partial-source-holdout-tinytcn",
    "status": "preview-only",
    "caveat": "From-scratch source-held-out stress test on 1 of 83 upstream shards. The train side is almost entirely synthetic while validation is mostly human English, so this measures a joint source/language/synthetic domain shift, not isolated causality or Hinglish quality."
  },
  "model": {
    "type": "tiny_tcn",
    "n_mels": 80,
    "channels": 128,
    "num_blocks": 6,
    "kernel_size": 5,
    "dilation_cycle": [1, 2, 4, 8, 16, 32],
    "attention_channels": 96,
    "head_hidden": 96,
    "dropout": 0.1,
    "auxiliary_fillers": true
  },
  "features": {
    "sample_rate": 16000,
    "n_fft": 400,
    "hop_length": 160,
    "win_length": 400,
    "n_mels": 80,
    "f_min": 0.0,
    "f_max": 8000.0,
    "normalize": false,
    "mel_scale": "htk",
    "log_scale": "whisper",
    "center": true,
    "drop_last_frame": true,
    "pad_side": "left",
    "max_seconds": 4.0
  },
  "data": {
    "train_source": "data/processed/partial-source-holdout.jsonl",
    "validation_source": "data/processed/partial-source-holdout.jsonl",
    "source_root": ".",
    "train_split": "train",
    "validation_split": "validation",
    "batch_size": 16,
    "num_workers": 0,
    "shuffle_buffer": 32,
    "revision": "e564e2ac567f774d1880aa1db6ce97afb8c519b7",
    "scope": "preview-only source-held-out stress: 3265 rows from train-00010-of-00083; official test remains sealed"
  },
  "loss": {
    "endpoint_weight": 1.0,
    "midfiller_weight": 0.15,
    "endfiller_weight": 0.15,
    "label_smoothing": 0.02,
    "endpoint_pos_weight": null
  },
  "training": {
    "seed": 17,
    "epochs": 3,
    "learning_rate": 0.0003,
    "weight_decay": 0.001,
    "gradient_accumulation_steps": 1,
    "max_gradient_norm": 1.0,
    "mixed_precision": false,
    "deterministic": true,
    "device": "cpu",
    "early_stopping_patience": 3,
    "min_improvement": 0.0001,
    "false_positive_rate_budget": 0.02,
    "selection_metric": "average_precision"
  }
}