gpt2-lang-ident

This model is a fine-tuned version of openai-community/gpt2 on sampled sentences from stanford-oval/ccnews and qanastek/EMEA-V3 datasets. It achieves the following results on the evaluation set:

• Loss: 0.1210
• Accuracy: 0.9721

Model description

This model is trained to predict the language of the input text.

Intended uses & limitations

This model can predict the following 90 languages:

[
    "af", "am", "ar", "as", "az", "ba", "be", "bg", "bn", "ca",
    "ceb", "ckb", "cs", "cy", "da", "de", "dv", "el", "en", "eo",
    "es", "et", "eu", "fa", "fi", "fr", "fy", "ga", "gd", "gl",
    "gu", "he", "hi", "hr", "hu", "hy", "id", "is", "it", "ja",
    "ka", "kk", "kn", "ku", "ky", "la", "lb", "lt", "lv", "mg",
    "mk", "ml", "mn", "mr", "mt", "my", "nds", "ne", "nl", "nn",
    "no", "or", "pa", "pl", "ps", "pt", "ro", "ru", "sah", "sd",
    "si", "sk", "sl", "sq", "sr", "sv", "sw", "ta", "te", "tg",
    "th", "tk", "tl", "tr", "tt", "ug", "uk", "ur", "vi", "yi"
]

How to use:

from transformers import (AutoModelForSequenceClassification, AutoTokenizer,
                          pipeline)

checkpoint = f"nie3e/gpt2-lang-ident"
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
tokenizer = AutoTokenizer.from_pretrained(checkpoint)

pipe = pipeline(
    task="text-classification",
    model=model,
    tokenizer=tokenizer,
    top_k=5
)

result = pipe("To jest model służący do identyfikacji języka!")
print(result)

[[{'label': 'pl', 'score': 0.9999653100967407}, {'label': 'sr', 'score': 1.5228776646836195e-05}, {'label': 'hr', 'score': 1.057955432770541e-05}, {'label': 'bn', 'score': 1.590750912328076e-06}, {'label': 'cs', 'score': 1.3942196801508544e-06}]]

Training and evaluation data

Training data ([lang]: count)

[bn]: 1947
[ar]: 1947
[vi]: 1947
[uk]: 1947
[kn]: 1947
[mr]: 1947
[id]: 1947
[te]: 1947
[no]: 1947
[ru]: 1947
[he]: 1947
[az]: 1947
[ca]: 1946
[fa]: 1946
[hi]: 1946
[th]: 1946
[tr]: 1946
[mk]: 1946
[ta]: 1945
[sq]: 1945
[ur]: 1942
[gu]: 1939
[ml]: 1936
[is]: 1738
[de]: 1543
[da]: 1521
[fi]: 1461
[el]: 1431
[nl]: 1424
[fr]: 1408
[cs]: 1401
[es]: 1397
[en]: 1394
[lt]: 1392
[hu]: 1379
[pt]: 1375
[lv]: 1373
[it]: 1360
[pl]: 1355
[sk]: 1355
[et]: 1348
[sl]: 1328
[sv]: 1300
[bg]: 1278
[mt]: 1234
[ro]: 1218
[kk]: 1179
[hy]: 1176
[or]: 1112
[pa]: 780
[sr]: 744
[as]: 735
[hr]: 722
[ne]: 626
[gl]: 566
[ckb]: 563
[ka]: 560
[ug]: 485
[ky]: 453
[eu]: 351
[ps]: 311
[tl]: 307
[fy]: 290
[mn]: 289
[si]: 244
[cy]: 214
[nn]: 212
[ku]: 195
[tg]: 176
[am]: 141
[tt]: 121
[ja]: 104
[lb]: 93
[tk]: 72
[be]: 64
[sw]: 45
[af]: 44
[my]: 40
[ceb]: 35
[la]: 33
[dv]: 20
[ba]: 19
[ga]: 19
[eo]: 19
[gd]: 16
[mg]: 15
[yi]: 14
[sah]: 14
[sd]: 11
[nds]: 11

Eval data ([lang]: count)

[te]: 195
[mk]: 195
[bn]: 195
[uk]: 195
[hi]: 195
[ar]: 195
[sq]: 195
[kn]: 195
[tr]: 195
[ca]: 195
[az]: 195
[fa]: 195
[ru]: 195
[mr]: 195
[id]: 195
[no]: 195
[vi]: 195
[th]: 195
[he]: 195
[gu]: 194
[ml]: 194
[ta]: 194
[ur]: 194
[is]: 174
[de]: 154
[da]: 152
[fi]: 146
[el]: 143
[nl]: 142
[fr]: 141
[es]: 140
[cs]: 140
[en]: 139
[lt]: 139
[hu]: 138
[lv]: 137
[pt]: 137
[it]: 136
[et]: 135
[pl]: 135
[sk]: 135
[sl]: 133
[sv]: 130
[bg]: 128
[mt]: 123
[ro]: 122
[hy]: 118
[kk]: 118
[or]: 111
[pa]: 78
[sr]: 74
[as]: 74
[hr]: 72
[ne]: 63
[gl]: 57
[ckb]: 56
[ka]: 56
[ug]: 49
[ky]: 45
[eu]: 35
[ps]: 31
[tl]: 31
[mn]: 29
[fy]: 29
[si]: 24
[nn]: 21
[cy]: 21
[ku]: 19
[tg]: 18
[am]: 14
[tt]: 12
[ja]: 10
[lb]: 9
[tk]: 7
[be]: 6
[my]: 4
[sw]: 4
[af]: 4
[ceb]: 3
[la]: 3
[ba]: 2
[dv]: 2
[eo]: 2
[gd]: 2
[ga]: 2
[mg]: 1
[sd]: 1
[nds]: 1
[yi]: 1
[sah]: 1

Training procedure

GPU: RTX 3090
Training time: 1h 53min

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Framework versions

• Transformers 4.36.2
• Pytorch 2.1.2+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0