XGLM-7.5B

XGLM-7.5B is a multilingual autoregressive language model (with 7.5 billion parameters) trained on a balanced corpus of a diverse set of languages totaling 500 billion sub-tokens. It was introduced in the paper Few-shot Learning with Multilingual Language Models 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.

Training Data Statistics

The training data statistics of XGLM-7.5B 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 released by the XGLM-7.5B 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.

import torch
import torch.nn.functional as F

from transformers import XGLMTokenizer, XGLMForCausalLM

tokenizer = XGLMTokenizer.from_pretrained("facebook/xglm-7.5B")
model = XGLMForCausalLM.from_pretrained("facebook/xglm-7.5B")

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