File size: 2,421 Bytes
8732fc1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
{
  "audio": {
    "sample_rate": 32000,
    "channels": "mono (average of channels)",
    "window_s": 10,
    "song_vector": "split into non-overlapping 10 s windows (keep a shorter last window if it is at least 2 s), embed each window, average the window vectors, L2-normalise",
    "front_end": "inside the ONNX model: pre-emphasis 0.97, STFT n_fft 1024 / hop 320 / Hann 800 centred, 128 Kaldi mel bands 0–15 kHz, log(x + 1e-5), (x + 4.5) / 5 (EfficientAT)",
    "onnx_input": "audio_32k float32 [batch, samples]",
    "onnx_output": "vibe_vector float32 [batch, 512], L2-normalised"
  },
  "text": {
    "model": "Qwen/Qwen3-Embedding-0.6B (fine-tuned)",
    "prompt": "Instruct: Given a description of how a piece of music sounds or feels, or of a moment it suits, retrieve music that matches it\nQuery:",
    "tokenizer": "left padding, truncation 256 tokens",
    "onnx_input": "input_ids, attention_mask int64 [batch, tokens] (prompt + query)",
    "onnx_output": "text_vector float32 [batch, 512], L2-normalised",
    "int8_method": "8-bit weights (block 32), int8 arithmetic (accuracy level 4) + 16-bit vocabulary",
    "int8_min_onnxruntime": "1.28",
    "int8_checks": {
      "cos_mean": 0.99976,
      "cos_min": 0.99919,
      "plugin_parity_min": 0.99979,
      "incompe_full": 0.7187,
      "ms_per_query_4thread": 55.8,
      "MB": 818.5
    },
    "version": "v1.2",
    "fine_tune": "Qwen3-Embedding-0.6B and the text layer trained to copy the combined teacher's 7B text vectors; translated training texts (es, de, fr, pt-BR, ja) map to their English text's vector",
    "languages": [
      "en",
      "es",
      "de",
      "fr",
      "pt",
      "ja"
    ]
  },
  "matryoshka": {
    "sizes": [
      128,
      256,
      512
    ],
    "how": "take the first N numbers and L2-normalise again"
  },
  "students": {
    "mn20_as": {
      "efficientat": "mn20_as",
      "efficientat_commit": "a425fdce92572e602a1d5634799bd9f1f2efa806",
      "params_m": 17.87,
      "best_epoch": 8,
      "v1_2": "C · trained on the teacher's song vectors only (epoch 3)"
    }
  },
  "default_student": "mn20_as",
  "selection": "epoch and size chosen on the dev halves of IncompeBench (nDCG@10, lenient) and Song Describer (R@10)",
  "teacher": "combined: LCO-Embedding-Omni-7B (LoRA) + e5-omni-7B, 512 numbers (teacher/combined)",
  "licence": "Apache-2.0 (EfficientAT code: MIT)",
  "version": "v1.2"
}