MIX1_ja-en_helsinki

This model is a fine-tuned version of Helsinki-NLP/opus-mt-ja-en on a combination of Visual Novel, Light Novel, and Subtitle data. A total of ~10MM lines of training data were used. It achieves the following results on the evaluation set:

• Loss: 1.7947
• Otaku Benchmark VN BLEU: 17.78
• Otaku Benchmark LN BLEU: 11.80
• Otaku Benchmark MANGA BLEU: 13.66

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0003
• train_batch_size: 64
• eval_batch_size: 64
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
2.7495	0.01	2000	2.5989
2.5415	0.03	4000	2.4746
2.4409	0.04	6000	2.4731
2.3743	0.05	8000	2.4012
2.3254	0.06	10000	2.3904
2.2857	0.08	12000	2.3649
2.2448	0.09	14000	2.3188
2.2158	0.1	16000	2.2975
2.193	0.11	18000	2.2756
2.1669	0.13	20000	2.2852
2.144	0.14	22000	2.2689
2.1222	0.15	24000	2.2721
2.1045	0.16	26000	2.2489
2.0885	0.18	28000	2.2359
2.0732	0.19	30000	2.2771
2.0584	0.2	32000	2.2582
2.0471	0.21	34000	2.2093
2.0369	0.23	36000	2.1768
2.0241	0.24	38000	2.1884
2.0196	0.25	40000	2.2025
2.004	0.27	42000	2.1507
1.9936	0.28	44000	2.1668
1.9869	0.29	46000	2.1432
1.9735	0.3	48000	2.1662
1.9651	0.32	50000	2.1824
1.9551	0.33	52000	2.1608
1.9485	0.34	54000	2.1322
1.9421	0.35	56000	2.1476
1.9303	0.37	58000	2.0994
1.9236	0.38	60000	2.1182
1.9183	0.39	62000	2.1305
1.9108	0.4	64000	2.1469
1.9051	0.42	66000	2.1414
1.9018	0.43	68000	2.1089
1.8959	0.44	70000	2.0908
1.886	0.46	72000	2.0968
1.8802	0.47	74000	2.0503
1.8713	0.48	76000	2.0542
1.8648	0.49	78000	2.0990
1.8599	0.51	80000	2.1112
1.8563	0.52	82000	2.1007
1.8541	0.53	84000	2.0849
1.845	0.54	86000	2.0831
1.8448	0.56	88000	2.0560
1.8342	0.57	90000	2.0349
1.8344	0.58	92000	2.0301
1.8291	0.59	94000	2.0300
1.819	0.61	96000	2.0378
1.8154	0.62	98000	2.0197
1.82	0.63	100000	2.0463
1.8081	0.64	102000	2.0077
1.8046	0.66	104000	2.0101
1.7978	0.67	106000	2.0150
1.7934	0.68	108000	2.0215
1.7904	0.7	110000	2.0278
1.7871	0.71	112000	2.0588
1.779	0.72	114000	2.0062
1.7784	0.73	116000	2.0300
1.7749	0.75	118000	1.9664
1.7691	0.76	120000	2.0033
1.7622	0.77	122000	1.9983
1.7587	0.78	124000	2.0030
1.755	0.8	126000	1.9955
1.7531	0.81	128000	1.9764
1.7439	0.82	130000	1.9942
1.7406	0.83	132000	2.0221
1.7385	0.85	134000	1.9835
1.7332	0.86	136000	1.9967
1.7332	0.87	138000	2.0247
1.7309	0.88	140000	1.9817
1.7248	0.9	142000	2.0063
1.7209	0.91	144000	1.9583
1.7154	0.92	146000	1.9779
1.7153	0.94	148000	1.9478
1.7094	0.95	150000	1.9706
1.7061	0.96	152000	1.9605
1.7017	0.97	154000	1.9447
1.6965	0.99	156000	1.9419
1.6929	1.0	158000	1.9589
1.6628	1.01	160000	1.9383
1.6535	1.02	162000	1.9487
1.6495	1.04	164000	1.9400
1.6516	1.05	166000	1.9353
1.6513	1.06	168000	1.9253
1.6518	1.07	170000	1.9132
1.6491	1.09	172000	1.9076
1.6453	1.1	174000	1.9192
1.6426	1.11	176000	1.9191
1.6353	1.13	178000	1.9367
1.6352	1.14	180000	1.9218
1.6304	1.15	182000	1.9305
1.6299	1.16	184000	1.9072
1.6263	1.18	186000	1.9211
1.6284	1.19	188000	1.9037
1.6237	1.2	190000	1.8951
1.6231	1.21	192000	1.8998
1.6184	1.23	194000	1.8960
1.6153	1.24	196000	1.8776
1.6122	1.25	198000	1.8747
1.6109	1.26	200000	1.8951
1.6072	1.28	202000	1.8705
1.6094	1.29	204000	1.8903
1.6063	1.3	206000	1.8660
1.599	1.31	208000	1.8696
1.5931	1.33	210000	1.8598
1.5943	1.34	212000	1.8760
1.5906	1.35	214000	1.8833
1.5858	1.37	216000	1.8645
1.5873	1.38	218000	1.8620
1.5842	1.39	220000	1.8632
1.5808	1.4	222000	1.8782
1.5756	1.42	224000	1.8627
1.5728	1.43	226000	1.8649
1.5709	1.44	228000	1.8735
1.5704	1.45	230000	1.8630
1.5659	1.47	232000	1.8598
1.5637	1.48	234000	1.8519
1.5628	1.49	236000	1.8569
1.5559	1.5	238000	1.8401
1.5532	1.52	240000	1.8528
1.557	1.53	242000	1.8637
1.5499	1.54	244000	1.8701
1.5476	1.55	246000	1.8423
1.5502	1.57	248000	1.8320
1.5469	1.58	250000	1.8542
1.5382	1.59	252000	1.8526
1.5396	1.61	254000	1.8537
1.528	1.62	256000	1.8248
1.532	1.63	258000	1.8322
1.5269	1.64	260000	1.8381
1.5269	1.66	262000	1.8389
1.5269	1.67	264000	1.8445
1.525	1.68	266000	1.8232
1.5175	1.69	268000	1.8561
1.5172	1.71	270000	1.8342
1.5174	1.72	272000	1.8167
1.5114	1.73	274000	1.8281
1.5094	1.74	276000	1.8164
1.5083	1.76	278000	1.8317
1.5047	1.77	280000	1.8207
1.5045	1.78	282000	1.8155
1.497	1.8	284000	1.8275
1.4996	1.81	286000	1.8152
1.497	1.82	288000	1.8137
1.4967	1.83	290000	1.8109
1.4936	1.85	292000	1.8037
1.4867	1.86	294000	1.7955
1.4859	1.87	296000	1.8181
1.4869	1.88	298000	1.7999
1.4811	1.9	300000	1.8062
1.4831	1.91	302000	1.8042
1.4791	1.92	304000	1.8020
1.4797	1.93	306000	1.7972
1.483	1.95	308000	1.8044
1.4748	1.96	310000	1.8036
1.4772	1.97	312000	1.7958
1.4708	1.98	314000	1.7967
1.4743	2.0	316000	1.7947

Framework versions

• Transformers 4.19.2
• Pytorch 1.11.0+cu113
• Datasets 2.2.2
• Tokenizers 0.12.1