File size: 8,366 Bytes
ee7140d
d3527c7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
603d4fa
 
ee7140d
603d4fa
 
92e3107
 
63fb4d6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
92e3107
 
 
48ecb9b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
7943153
 
 
 
 
 
 
 
 
 
 
 
 
 
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
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
---
language:
- multilingual
- en
- ru
- zh
- de
- es
- fr
- ja
- it
- pt
- el
- ko
- fi
- id
- tr
- ar
- vi
- th
- bg
- ca
- hi
- et
- bn
- ta
- ur
- sw
- te
- eu
- my
- ht
- qu
license: mit
thumbnail: https://huggingface.co/front/thumbnails/facebook.png
inference: false
---

# XGLM-564M

XGLM-564M is a multilingual autoregressive language model (with 564 million parameters) trained on a balanced corpus of a diverse set of 30 languages totaling 500 billion sub-tokens. It was introduced in the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin\*, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li\* (\*Equal Contribution). The original implementation was released in [this repository](https://github.com/pytorch/fairseq/tree/main/examples/xglm).

## Training Data Statistics

The training data statistics of XGLM-564M is shown in the table below.

| ISO-639-1| family          | name                    |  # tokens    |       ratio |   ratio w/ lowRes upsampling |
|:--------|:-----------------|:------------------------|-------------:|------------:|-------------:|
| en      | Indo-European    | English                 | 803526736124 | 0.489906    |       0.3259 |
| ru      | Indo-European    | Russian                 | 147791898098 | 0.0901079   |       0.0602 |
| zh      | Sino-Tibetan     | Chinese                 | 132770494630 | 0.0809494   |       0.0483 |
| de      | Indo-European    | German                  |  89223707856 | 0.0543992   |       0.0363 |
| es      | Indo-European    | Spanish                 |  87303083105 | 0.0532282   |       0.0353 |
| fr      | Indo-European    | French                  |  77419639775 | 0.0472023   |       0.0313 |
| ja      | Japonic          | Japanese                |  66054364513 | 0.040273    |       0.0269 |
| it      | Indo-European    | Italian                 |  41930465338 | 0.0255648   |       0.0171 |
| pt      | Indo-European    | Portuguese              |  36586032444 | 0.0223063   |       0.0297 |
| el      | Indo-European    | Greek (modern)          |  28762166159 | 0.0175361   |       0.0233 |
| ko      | Koreanic         | Korean                  |  20002244535 | 0.0121953   |       0.0811 |
| fi      | Uralic           | Finnish                 |  16804309722 | 0.0102455   |       0.0681 |
| id      | Austronesian     | Indonesian              |  15423541953 | 0.00940365  |       0.0125 |
| tr      | Turkic           | Turkish                 |  12413166065 | 0.00756824  |       0.0101 |
| ar      | Afro-Asiatic     | Arabic                  |  12248607345 | 0.00746791  |       0.0099 |
| vi      | Austroasiatic    | Vietnamese              |  11199121869 | 0.00682804  |       0.0091 |
| th      | Tai–Kadai        | Thai                    |  10842172807 | 0.00661041  |       0.044  |
| bg      | Indo-European    | Bulgarian               |   9703797869 | 0.00591635  |       0.0393 |
| ca      | Indo-European    | Catalan                 |   7075834775 | 0.0043141   |       0.0287 |
| hi      | Indo-European    | Hindi                   |   3448390110 | 0.00210246  |       0.014  |
| et      | Uralic           | Estonian                |   3286873851 | 0.00200399  |       0.0133 |
| bn      | Indo-European    | Bengali, Bangla         |   1627447450 | 0.000992245 |       0.0066 |
| ta      | Dravidian        | Tamil                   |   1476973397 | 0.000900502 |       0.006  |
| ur      | Indo-European    | Urdu                    |   1351891969 | 0.000824241 |       0.0055 |
| sw      | Niger–Congo      | Swahili                 |    907516139 | 0.000553307 |       0.0037 |
| te      | Dravidian        | Telugu                  |    689316485 | 0.000420272 |       0.0028 |
| eu      | Language isolate | Basque                  |    105304423 | 6.42035e-05 |       0.0043 |
| my      | Sino-Tibetan     | Burmese                 |    101358331 | 6.17976e-05 |       0.003  |
| ht      | Creole           | Haitian, Haitian Creole |     86584697 | 5.27902e-05 |       0.0035 |
| qu      | Quechuan         | Quechua                 |      3236108 | 1.97304e-06 |       0.0001 |

## Model card

For intended usage of the model, please refer to the [model card](https://github.com/pytorch/fairseq/blob/main/examples/xglm/model_card.md) released by the XGLM-564M development team. 

## Example (COPA)
The following snippet shows how to evaluate our models (GPT-3 style, zero-shot) on the Choice of Plausible Alternatives (COPA) task, using examples in English, Chinese and Hindi.

```python
import torch
import torch.nn.functional as F

from transformers import XGLMTokenizer, XGLMForCausalLM

tokenizer = XGLMTokenizer.from_pretrained("facebook/xglm-564M")
model = XGLMForCausalLM.from_pretrained("facebook/xglm-564M")

data_samples = {
    'en': [
        {
            "premise": "I wanted to conserve energy.",
            "choice1": "I swept the floor in the unoccupied room.",
            "choice2": "I shut off the light in the unoccupied room.",
            "question": "effect",
            "label": "1"
        },
        {
            "premise": "The flame on the candle went out.",
            "choice1": "I blew on the wick.",
            "choice2": "I put a match to the wick.",
            "question": "cause",
            "label": "0"
        }
    ],
    'zh': [
        {
            "premise": "我想节约能源。",
            "choice1": "我在空着的房间里扫了地板。",
            "choice2": "我把空房间里的灯关了。",
            "question": "effect",
            "label": "1"
        },
        {
            "premise": "蜡烛上的火焰熄灭了。",
            "choice1": "我吹灭了灯芯。",
            "choice2": "我把一根火柴放在灯芯上。",
            "question": "cause",
            "label": "0"
        }
    ],
    'hi': [
        {
            "premise": "M te vle konsève enèji.",
            "choice1": "Mwen te fin baleye chanm lib la.",
            "choice2": "Mwen te femen limyè nan chanm lib la.",
            "question": "effect",
            "label": "1"
        },
        {
            "premise": "Flam bouji a te etenn.",
            "choice1": "Mwen te soufle bouji a.",
            "choice2": "Mwen te limen mèch bouji a.",
            "question": "cause",
            "label": "0"
        }
    ]
}

def get_logprobs(prompt):
    inputs = tokenizer(prompt, return_tensors="pt")
    input_ids, output_ids = inputs["input_ids"], inputs["input_ids"][:, 1:]
    outputs = model(**inputs, labels=input_ids)
    logits = outputs.logits
    logprobs = torch.gather(F.log_softmax(logits, dim=2), 2, output_ids.unsqueeze(2))
    return logprobs

# Zero-shot evaluation for the Choice of Plausible Alternatives (COPA) task.
# A return value of 0 indicates that the first alternative is more plausible,
# while 1 indicates that the second alternative is more plausible.
def COPA_eval(prompt, alternative1, alternative2):
    lprob1 = get_logprobs(prompt + "\n" + alternative1).sum()
    lprob2 = get_logprobs(prompt + "\n" + alternative2).sum()
    return 0 if lprob1 > lprob2 else 1

for lang in data_samples_long:
    for idx, example in enumerate(data_samples_long[lang]):
        predict = COPA_eval(example["premise"], example["choice1"], example["choice2"])
        print(f'{lang}-{idx}', predict, example['label'])
        
# en-0 1 1
# en-1 0 0
# zh-0 1 1
# zh-1 0 0
# hi-0 1 1
# hi-1 0 0
```
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_facebook__xglm-564M)

| Metric                | Value                     |
|-----------------------|---------------------------|
| Avg.                  | 26.19   |
| ARC (25-shot)         | 24.57          |
| HellaSwag (10-shot)   | 34.64    |
| MMLU (5-shot)         | 25.18         |
| TruthfulQA (0-shot)   | 40.43   |
| Winogrande (5-shot)   | 52.25   |
| GSM8K (5-shot)        | 0.23        |
| DROP (3-shot)         | 6.04         |