---
datasets:
- code_search_net
widget:
- text: "def <mask> ( a, b ) : if a > b : return a else return b</s>return the maximum value"
- text: "def <mask> ( a, b ) : if a > b : return a else return b"
---

# Model Architecture

This model follows the distilroberta-base architecture. Futhermore, this model was initialized with the checkpoint of distilroberta-base.

# Pre-training phase

This model was pre-trained with the MLM objective (`mlm_probability=0.15`).

During this phase, the inputs had two formats. One is the following:
$$\left[[CLS], t_1, \dots, t_n, [SEP], w_1, \dots, w_m\right[EOS]]$$
where $t_1, \dots, t_n$ are the code tokens and $w_1, \dots, w_m$ are the natural language description tokens. More concretely, this is the snippet that tokenizes the input:
```python
def tokenize_function_bimodal(examples, tokenizer, max_len):
    codes = [' '.join(example) for example in examples['func_code_tokens']]
    nls = [' '.join(example) for example in examples['func_documentation_tokens']]
    pairs = [[c, nl] for c, nl in zip(codes, nls)]
    return tokenizer(pairs, max_length=max_len, padding="max_length", truncation=True)
```
The other format is:
$$\left[[CLS], t_1, \dots, t_n \right[EOS]]$$
where $t_1, \dots, t_n$ are the code tokens. More concretely, this is the snippet that tokenizes the input:

```python
def tokenize_function_unimodal(examples, tokenizer, max_len, tokens_column):
    codes = [' '.join(example) for example in examples[tokens_column]]
    return tokenizer(codes, max_length=max_len, padding="max_length", truncation=True)
```


# Training details

- Max length: 512
- Effective batch size: 64
- Total steps: 140000
- Learning rate: 5e-4

# Usage

```python
model = AutoModelForMaskedLM.from_pretrained('antolin/distilroberta-base-csn-python-unimodal-bimodal')
tokenizer = AutoTokenizer.from_pretrained('antolin/distilroberta-base-csn-python-unimodal-bimodal')
mask_filler = pipeline("fill-mask", model=model, tokenizer=tokenizer)
code_tokens = ["def", "<mask>", "(", "a", ",", "b", ")", ":", "if", "a", ">", "b", ":", "return", "a", "else", "return", "b"]
nl_tokens = ["return", "the", "maximum", "value"]
input_text = ' '.join(code_tokens) + tokenizer.sep_token + ' '.join(nl_tokens)
pprint(mask_filler(input_text, top_k=5))
```
```shell
[{'score': 0.7177600860595703,
  'sequence': 'def maximum ( a, b ) : if a > b : return a else return breturn '
              'the maximum value',
  'token': 4532,
  'token_str': ' maximum'},
 {'score': 0.22075247764587402,
  'sequence': 'def max ( a, b ) : if a > b : return a else return breturn the '
              'maximum value',
  'token': 19220,
  'token_str': ' max'},
 {'score': 0.015111264772713184,
  'sequence': 'def minimum ( a, b ) : if a > b : return a else return breturn '
              'the maximum value',
  'token': 3527,
  'token_str': ' minimum'},
 {'score': 0.007394665852189064,
  'sequence': 'def min ( a, b ) : if a > b : return a else return breturn the '
              'maximum value',
  'token': 5251,
  'token_str': ' min'},
 {'score': 0.004020793363451958,
  'sequence': 'def length ( a, b ) : if a > b : return a else return breturn '
              'the maximum value',
  'token': 5933,
  'token_str': ' length'}]
```