Hugging Face's logo

LIAMF-USP
/
aristo-roberta

Multiple Choice
Transformers
PyTorch
TensorFlow
JAX
roberta
Inference Endpoints
Model card Files Community
1
aristo-roberta
File size: 4,167 Bytes 
18f8c91
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
 
---
language: "english"
tags:
license: "mit"
datasets:
- race
- arc
metrics:
- accuracy
---

# Roberta Large Fine Tuned on RACE

## Model description

This model follows the implementation by Allen AI team about [Aristo Roberta V7 Model](https://leaderboard.allenai.org/arc/submission/blcotvl7rrltlue6bsv0) given in [ARC Challenge](https://leaderboard.allenai.org/arc/submissions/public)

#### How to use

```python

import datasets
from transformers import RobertaTokenizer
from transformers import  RobertaForMultipleChoice

tokenizer = RobertaTokenizer.from_pretrained(
"LIAMF-USP/aristo-roberta")
model = RobertaForMultipleChoice.from_pretrained(
"LIAMF-USP/aristo-roberta")
dataset = datasets.load_dataset(
    "arc",,
    split=["train", "validation", "test"],
)
training_examples = dataset[0]
evaluation_examples = dataset[1]
test_examples = dataset[2]

example=training_examples[0] 
example_id = example["example_id"]
question = example["question"]
label_example = example["answer"]
options = example["options"]
if label_example in ["A", "B", "C", "D", "E"]:
    label_map = {label: i for i, label in enumerate(
					["A", "B", "C", "D", "E"])}
elif label_example in ["1", "2", "3", "4", "5"]:
    label_map = {label: i for i, label in enumerate(
					["1", "2", "3", "4", "5"])}
else:
    print(f"{label_example} not found")
while len(options) < 5:
    empty_option = {}
    empty_option['option_context'] = ''
    empty_option['option_text'] = ''
    options.append(empty_option)
choices_inputs = []
for ending_idx, option in enumerate(options):
    ending = option["option_text"]
    context = option["option_context"]
    if question.find("_") != -1:
        # fill in the banks questions
        question_option = question.replace("_", ending)
    else:
        question_option = question + " " + ending
    
    inputs = tokenizer(
        context,
        question_option,
        add_special_tokens=True,
        max_length=MAX_SEQ_LENGTH,
        padding="max_length",
        truncation=True,
        return_overflowing_tokens=False,
    )
    
    if "num_truncated_tokens" in inputs and inputs["num_truncated_tokens"] > 0:
        logging.warning(f"Question: {example_id} with option {ending_idx} was truncated")
    choices_inputs.append(inputs)
label = label_map[label_example]
input_ids = [x["input_ids"] for x in choices_inputs]
attention_mask = (
    [x["attention_mask"] for x in choices_inputs]
     # as the senteces follow the same structure, just one of them is
     # necessary to check
    if "attention_mask" in choices_inputs[0]
    else None
)
example_encoded = {
    "example_id": example_id,
    "input_ids": input_ids,
    "attention_mask": attention_mask,
    "token_type_ids": token_type_ids,
    "label": label

}
output = model(**example_encoded)
```


## Training data

the Training data was the same as proposed [here](https://leaderboard.allenai.org/arc/submission/blcotvl7rrltlue6bsv0) 

The only diferrence was the hypeparameters of RACE fine tuned model, which were reported [here](https://huggingface.co/LIAMF-USP/roberta-large-finetuned-race#eval-results)

## Training procedure

It was necessary to preprocess the data with a method that is exemplified for a single instance in the _How to use_ section. The used hyperparameters were the following:

| Hyperparameter | Value |
|:----:|:----:|
| adam_beta1                  | 0.9      |
| adam_beta2                  | 0.98     |
| adam_epsilon                | 1.000e-8 |
| eval_batch_size             | 16       |
| train_batch_size            | 4        |
| fp16                        | True     |
| gradient_accumulation_steps | 4       |
| learning_rate               | 0.00001  |
| warmup_steps                | 67     |
| max_length                  | 256      |
| epochs                  |   2      |

The other parameters were the default ones from [Trainer](https://huggingface.co/transformers/main_classes/trainer.html) and [Trainer Arguments](https://huggingface.co/transformers/main_classes/trainer.html#trainingarguments)

## Eval results:
| Dataset Acc | Challenge Test |
|:----:|:----:|
|      | 64.249 |

**The model was trained with a TITAN RTX**