AlexCode2003 commited on
Commit
00d4d9e
·
verified ·
1 Parent(s): 448ef3c

Прогон v2: эмбеддинги, дообучение, варианты квантизации

Browse files
v2/dumps/encoder_fp32.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:14c674307276b5921ee98565196369551e4e2f8c04b1bcad9f19fa3164994e63
3
+ size 2701647
v2/dumps/encoder_int8_emb_only.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a5a38c5efe400e23089236797d9b4afb7a214b6dd22b4f10e317f18352296687
3
+ size 2701567
v2/dumps/encoder_int8_full.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:724e008b30551a3b38a956c4937b87cd6a12499f424e14f48198ac0e1afaea51
3
+ size 2699856
v2/dumps/encoder_int8_full_perchannel.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:69be1b2a7b7e5d99369207e71dc4b3acfbb8be0bc40f82a21f8bb17ce1bc7d33
3
+ size 2701333
v2/dumps/encoder_torch_fp32.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0d69db4da6cf021232502222f9fb30c47cba48394a1d93df0eea9fb11a188607
3
+ size 2701624
v2/dumps/finetuned_torch_fp32.npz ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d91e45799c47ce0de0fa4a1ad866cfb1949a8ab26c22592fff85941ff36aeec6
3
+ size 632625
v2/dumps/texts.json ADDED
The diff for this file is too large to render. See raw diff
 
v2/encoder_int8_emb_only.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2ebc9fc66ba19f62b2a799d0ff47c1a15d5d086106f954d5d4a42d96e54d09a0
3
+ size 181711482
v2/encoder_int8_full.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7a0e292f503a41525ecc675f40f0cd597c272fd6f98d8dc10a932ab21a92ce27
3
+ size 118108055
v2/encoder_int8_full_perchannel.onnx ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7dcf738ddfe05812a3ee198ab9626010673977bf7cf709a4a0f5a063318f9241
3
+ size 118315775
v2/metrics_v2.json ADDED
The diff for this file is too large to render. See raw diff
 
v2/report_v2.md ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # multilingual-e5-small, прогон v2
2
+
3
+ Значения – доля верных ответов, в скобках 95% доверительный интервал Уилсона.
4
+
5
+ ## Замороженный e5 против корректного дообучения (PyTorch fp32)
6
+
7
+ | Модель | CV acc | Test top-1 | macro-F1 | Порог | Test с порогом | OOD отклонено | Внешний тест | Сценарии |
8
+ |---|---|---|---|---|---|---|---|---|
9
+ | `e5_frozen_logreg` | 0.803 | 0.881 [0.840; 0.912] | 0.884 | 0.731 | 0.695 [0.642; 0.743] | 0.769 [0.664; 0.849] | 0.627 [0.559; 0.690] | 0.846 [0.739; 0.914] |
10
+ | `e5_finetuned_meanpool` | 0.749 | 0.811 [0.765; 0.851] | 0.806 | 0.379 | 0.701 [0.649; 0.749] | 0.628 [0.517; 0.727] | 0.627 [0.559; 0.690] | 0.862 [0.757; 0.925] |
11
+
12
+ ## Варианты ONNX: размер, память, задержка (CPU, 1 поток), качество
13
+
14
+ | Вариант | Размер, МБ | Память сессии, МБ | p50, мс | p95, мс | Совпадение top-1 с fp32 | Test top-1 | Сценарии с порогом |
15
+ |---|---|---|---|---|---|---|---|
16
+ | `encoder/fp32` | 448 | 516 | 12.4 | 22.7 | 1.000 | 0.881 [0.840; 0.912] | 0.846 [0.739; 0.914] |
17
+ | `encoder/int8_full` | 113 | 151 | 7.2 | 14.6 | 0.953 | 0.868 [0.826; 0.901] | 0.769 [0.654; 0.855] |
18
+ | `encoder/int8_full_perchannel` | 113 | 147 | 8.1 | 16.7 | 0.968 | 0.884 [0.844; 0.914] | 0.815 [0.704; 0.891] |
19
+ | `encoder/int8_emb_only` | 173 | 242 | 13.5 | 22.6 | 0.997 | 0.881 [0.840; 0.912] | 0.831 [0.722; 0.903] |
v2/tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:255f5e32cb320b4dc2b31e0f203832fecafe87ca4cc87c3c41eb8ccae73f441a
3
+ size 17082987
v2/tokenizer_config.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": true,
3
+ "backend": "tokenizers",
4
+ "bos_token": "<s>",
5
+ "clean_up_tokenization_spaces": true,
6
+ "cls_token": "<s>",
7
+ "eos_token": "</s>",
8
+ "is_local": false,
9
+ "local_files_only": false,
10
+ "mask_token": "<mask>",
11
+ "model_max_length": 512,
12
+ "pad_token": "<pad>",
13
+ "sep_token": "</s>",
14
+ "sp_model_kwargs": {},
15
+ "tokenizer_class": "XLMRobertaTokenizer",
16
+ "unk_token": "<unk>"
17
+ }