Initial commit

README.md

@@ -1,3 +1,59 @@
 ---
-license: cc-by-nc-sa-4.0
+language:
+- en
+- de
+tags:
+- translation
+- opus-mt-tc
+license: cc-by-4.0
+model-index:
+- name: opus-mt-tc-base-en-de
+  results:
+  - task:
+      name: Translation eng-deu
+      type: translation
+      args: eng-deu
+    dataset:
+      name: tatoeba-test-v2021-08-07
+      type: tatoeba_mt
+      args: eng-deu
+    metrics:
+    - name: BLEU
+      type: bleu
+      value: 43.7
 ---
+
+# Opus Tatoeba English-German
+
+*This model was obtained by running the script [convert_marian_to_pytorch.py](https://github.com/huggingface/transformers/blob/master/src/transformers/models/marian/convert_marian_to_pytorch.py). The original models were trained by [Jörg Tiedemann](https://blogs.helsinki.fi/tiedeman/) using the [MarianNMT](https://marian-nmt.github.io/) library. See all available `MarianMTModel` models on the profile of the [Helsinki NLP](https://huggingface.co/Helsinki-NLP) group.*
+
+* dataset: opusTCv20210807+bt
+* model: transformer-big
+* source language(s): eng
+* target language(s): deu
+* raw source language(s): eng
+* raw target language(s): deu
+* model: transformer-big
+* pre-processing: normalization + SentencePiece (spm32k,spm32k)
+* download: [opusTCv20210807+bt-2021-12-08.zip](https://object.pouta.csc.fi/Tatoeba-MT-models/eng-deu/opusTCv20210807+bt-2021-12-08.zip)
+* test set translations: [opusTCv20210807+bt-2021-12-08.test.txt](https://object.pouta.csc.fi/Tatoeba-MT-models/eng-deu/opusTCv20210807+bt-2021-12-08.test.txt)
+* test set scores: [opusTCv20210807+bt-2021-12-08.eval.txt](https://object.pouta.csc.fi/Tatoeba-MT-models/eng-deu/opusTCv20210807+bt-2021-12-08.eval.txt)
+
+## Benchmarks
+
+| testset | BLEU  | chr-F | #sent | #words | BP |
+|---------|-------|-------|-------|--------|----|
+| newssyscomb2009.eng-deu 	| 24.3 	| 0.5462 	| 502 	| 11271 	| 0.993 |
+| news-test2008.eng-deu 	| 24.7 	| 0.5412 	| 2051 	| 47427 	| 1.000 |
+| newstest2009.eng-deu 	| 23.6 	| 0.5385 	| 2525 	| 62816 	| 0.999 |
+| newstest2010.eng-deu 	| 26.9 	| 0.5589 	| 2489 	| 61511 	| 0.966 |
+| newstest2011.eng-deu 	| 24.1 	| 0.5364 	| 3003 	| 72981 	| 0.990 |
+| newstest2012.eng-deu 	| 24.6 	| 0.5375 	| 3003 	| 72886 	| 0.972 |
+| newstest2013.eng-deu 	| 28.3 	| 0.5636 	| 3000 	| 63737 	| 0.988 |
+| newstest2014-deen.eng-deu 	| 30.9 	| 0.6084 	| 3003 	| 62964 	| 1.000 |
+| newstest2015-ende.eng-deu 	| 33.2 	| 0.6106 	| 2169 	| 44260 	| 1.000 |
+| newstest2016-ende.eng-deu 	| 39.8 	| 0.6595 	| 2999 	| 62670 	| 0.993 |
+| newstest2017-ende.eng-deu 	| 32.0 	| 0.6047 	| 3004 	| 61291 	| 1.000 |
+| newstest2018-ende.eng-deu 	| 48.8 	| 0.7146 	| 2998 	| 64276 	| 1.000 |
+| newstest2019-ende.eng-deu 	| 45.0 	| 0.6821 	| 1997 	| 48969 	| 0.995 |
+| Tatoeba-test-v2021-08-07.eng-deu 	| 43.7 	| 0.6442 	| 10000 	| 85728 	| 1.000 |

config.json

@@ -0,0 +1,45 @@
+{
+  "activation_dropout": 0.0,
+  "activation_function": "relu",
+  "architectures": [
+    "MarianMTModel"
+  ],
+  "attention_dropout": 0.0,
+  "bad_words_ids": [
+    [
+      65000
+    ]
+  ],
+  "bos_token_id": 0,
+  "classifier_dropout": 0.0,
+  "d_model": 1024,
+  "decoder_attention_heads": 16,
+  "decoder_ffn_dim": 4096,
+  "decoder_layerdrop": 0.0,
+  "decoder_layers": 6,
+  "decoder_start_token_id": 65000,
+  "decoder_vocab_size": 65001,
+  "dropout": 0.1,
+  "encoder_attention_heads": 16,
+  "encoder_ffn_dim": 4096,
+  "encoder_layerdrop": 0.0,
+  "encoder_layers": 6,
+  "eos_token_id": 0,
+  "forced_eos_token_id": 0,
+  "init_std": 0.02,
+  "is_encoder_decoder": true,
+  "max_length": 512,
+  "max_position_embeddings": 1024,
+  "model_type": "marian",
+  "normalize_embedding": false,
+  "num_beams": 6,
+  "num_hidden_layers": 6,
+  "pad_token_id": 65000,
+  "scale_embedding": true,
+  "share_encoder_decoder_embeddings": true,
+  "static_position_embeddings": true,
+  "torch_dtype": "float16",
+  "transformers_version": "4.21.0.dev0",
+  "use_cache": true,
+  "vocab_size": 65001
+}

opusTCv20210807+bt.spm32k-spm32k.transformer-big.valid1.log

@@ -0,0 +1,56 @@
+[2021-12-01 16:52:50] [valid] Ep. 1 : Up. 10000 : perplexity : 4.62405 : new best
+[2021-12-01 18:42:19] [valid] Ep. 1 : Up. 20000 : perplexity : 3.17074 : new best
+[2021-12-01 20:31:43] [valid] Ep. 1 : Up. 30000 : perplexity : 2.89787 : new best
+[2021-12-01 22:21:07] [valid] Ep. 1 : Up. 40000 : perplexity : 2.76201 : new best
+[2021-12-02 00:10:29] [valid] Ep. 1 : Up. 50000 : perplexity : 2.68948 : new best
+[2021-12-02 01:59:57] [valid] Ep. 1 : Up. 60000 : perplexity : 2.64063 : new best
+[2021-12-02 03:49:30] [valid] Ep. 1 : Up. 70000 : perplexity : 2.60022 : new best
+[2021-12-02 05:38:44] [valid] Ep. 1 : Up. 80000 : perplexity : 2.56641 : new best
+[2021-12-02 07:28:03] [valid] Ep. 1 : Up. 90000 : perplexity : 2.53931 : new best
+[2021-12-02 09:17:19] [valid] Ep. 1 : Up. 100000 : perplexity : 2.51865 : new best
+[2021-12-02 11:06:38] [valid] Ep. 1 : Up. 110000 : perplexity : 2.50066 : new best
+[2021-12-02 12:55:59] [valid] Ep. 1 : Up. 120000 : perplexity : 2.48827 : new best
+[2021-12-02 14:45:20] [valid] Ep. 1 : Up. 130000 : perplexity : 2.47352 : new best
+[2021-12-02 16:34:59] [valid] Ep. 1 : Up. 140000 : perplexity : 2.46277 : new best
+[2021-12-02 18:24:13] [valid] Ep. 1 : Up. 150000 : perplexity : 2.4536 : new best
+[2021-12-02 20:13:45] [valid] Ep. 1 : Up. 160000 : perplexity : 2.44542 : new best
+[2021-12-02 22:03:11] [valid] Ep. 1 : Up. 170000 : perplexity : 2.43751 : new best
+[2021-12-02 23:52:31] [valid] Ep. 1 : Up. 180000 : perplexity : 2.42974 : new best
+[2021-12-03 01:41:53] [valid] Ep. 1 : Up. 190000 : perplexity : 2.42317 : new best
+[2021-12-03 17:40:24] [valid] Ep. 1 : Up. 200000 : perplexity : 2.41677 : new best
+[2021-12-03 19:30:07] [valid] Ep. 1 : Up. 210000 : perplexity : 2.41299 : new best
+[2021-12-03 21:19:37] [valid] Ep. 1 : Up. 220000 : perplexity : 2.40513 : new best
+[2021-12-03 23:09:03] [valid] Ep. 1 : Up. 230000 : perplexity : 2.39959 : new best
+[2021-12-04 00:58:20] [valid] Ep. 1 : Up. 240000 : perplexity : 2.39771 : new best
+[2021-12-04 02:47:43] [valid] Ep. 1 : Up. 250000 : perplexity : 2.39285 : new best
+[2021-12-04 04:37:02] [valid] Ep. 1 : Up. 260000 : perplexity : 2.38914 : new best
+[2021-12-04 06:26:43] [valid] Ep. 1 : Up. 270000 : perplexity : 2.38545 : new best
+[2021-12-04 08:15:58] [valid] Ep. 1 : Up. 280000 : perplexity : 2.38241 : new best
+[2021-12-04 10:05:13] [valid] Ep. 1 : Up. 290000 : perplexity : 2.37948 : new best
+[2021-12-04 11:54:23] [valid] Ep. 1 : Up. 300000 : perplexity : 2.37573 : new best
+[2021-12-04 13:43:34] [valid] Ep. 1 : Up. 310000 : perplexity : 2.37134 : new best
+[2021-12-04 15:32:47] [valid] Ep. 1 : Up. 320000 : perplexity : 2.36864 : new best
+[2021-12-04 17:22:00] [valid] Ep. 1 : Up. 330000 : perplexity : 2.36621 : new best
+[2021-12-04 19:11:32] [valid] Ep. 1 : Up. 340000 : perplexity : 2.36401 : new best
+[2021-12-04 21:00:49] [valid] Ep. 1 : Up. 350000 : perplexity : 2.36358 : new best
+[2021-12-04 22:50:04] [valid] Ep. 1 : Up. 360000 : perplexity : 2.36217 : new best
+[2021-12-05 00:39:08] [valid] Ep. 1 : Up. 370000 : perplexity : 2.36086 : new best
+[2021-12-05 02:28:18] [valid] Ep. 1 : Up. 380000 : perplexity : 2.35793 : new best
+[2021-12-05 16:54:50] [valid] Ep. 1 : Up. 390000 : perplexity : 2.35314 : new best
+[2021-12-05 18:43:23] [valid] Ep. 1 : Up. 400000 : perplexity : 2.35098 : new best
+[2021-12-05 20:32:10] [valid] Ep. 1 : Up. 410000 : perplexity : 2.35121 : stalled 1 times (last best: 2.35098)
+[2021-12-05 22:21:06] [valid] Ep. 1 : Up. 420000 : perplexity : 2.34825 : new best
+[2021-12-06 00:09:43] [valid] Ep. 1 : Up. 430000 : perplexity : 2.34765 : new best
+[2021-12-06 01:58:38] [valid] Ep. 1 : Up. 440000 : perplexity : 2.3442 : new best
+[2021-12-06 03:47:17] [valid] Ep. 1 : Up. 450000 : perplexity : 2.34198 : new best
+[2021-12-06 05:36:09] [valid] Ep. 1 : Up. 460000 : perplexity : 2.34093 : new best
+[2021-12-06 07:24:44] [valid] Ep. 1 : Up. 470000 : perplexity : 2.33995 : new best
+[2021-12-06 09:13:39] [valid] Ep. 1 : Up. 480000 : perplexity : 2.33989 : new best
+[2021-12-06 11:02:16] [valid] Ep. 1 : Up. 490000 : perplexity : 2.3377 : new best
+[2021-12-06 12:51:03] [valid] Ep. 1 : Up. 500000 : perplexity : 2.33755 : new best
+[2021-12-06 14:39:34] [valid] Ep. 1 : Up. 510000 : perplexity : 2.33954 : stalled 1 times (last best: 2.33755)
+[2021-12-06 16:27:44] [valid] Ep. 1 : Up. 520000 : perplexity : 2.34083 : stalled 2 times (last best: 2.33755)
+[2021-12-06 18:16:02] [valid] Ep. 1 : Up. 530000 : perplexity : 2.34335 : stalled 3 times (last best: 2.33755)
+[2021-12-06 20:04:29] [valid] Ep. 1 : Up. 540000 : perplexity : 2.3448 : stalled 4 times (last best: 2.33755)
+[2021-12-06 21:53:00] [valid] Ep. 1 : Up. 550000 : perplexity : 2.34299 : stalled 5 times (last best: 2.33755)
+[2021-12-06 23:41:33] [valid] Ep. 1 : Up. 560000 : perplexity : 2.34046 : stalled 6 times (last best: 2.33755)

pytorch_model.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:94ff3bc13fb7c6e175ced0ce952d1ec47f8d836fe70045c0f2e6cbc78e92bd53
+size 619169923

special_tokens_map.json

@@ -0,0 +1,5 @@
+{
+  "eos_token": "</s>",
+  "pad_token": "<pad>",
+  "unk_token": "<unk>"
+}

tokenizer_config.json

@@ -0,0 +1,14 @@
+{
+  "eos_token": "</s>",
+  "model_max_length": 512,
+  "name_or_path": "models/opus-mt-tc-big-en-de",
+  "pad_token": "<pad>",
+  "separate_vocabs": false,
+  "source_lang": "opus-mt-tc-big-en",
+  "sp_model_kwargs": {},
+  "special_tokens_map_file": null,
+  "target_lang": "de",
+  "tokenizer_class": "MarianTokenizer",
+  "tokenizer_file": null, 
+  "unk_token": "<unk>"
+}