File size: 2,801 Bytes
a6bb176
 
e77286d
 
 
 
 
 
 
139a5db
a6bb176
139a5db
 
ad7a79a
e77286d
 
 
139a5db
 
 
 
 
 
e77286d
 
 
 
 
 
 
 
 
 
 
a4d4ad6
 
e77286d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
71fb641
 
e77286d
 
 
 
139a5db
 
e77286d
 
 
 
 
 
 
 
 
 
 
 
 
 
a4d4ad6
e77286d
 
 
 
 
 
 
 
a4d4ad6
 
 
 
 
 
139a5db
e77286d
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
---
license: apache-2.0
language:
- en
tags:
- code
- knowledge extraction
- tiny
- small
- C
---
## Model info

A model that can **extract the knowledge points** from the given **C language code**. 

The base model is [pythia-70m](https://huggingface.co/EleutherAI/pythia-70m). This model was fine-tuned with 10 epochs using [Q-Lora](https://github.com/artidoro/qlora) method on my own training set.



## How to use

### quick start

A usage example is as follows, first import the model and prepare the code:

```python
from transformers import GPTNeoXForCausalLM, AutoTokenizer

model_name_or_path = 'Mxode/Pythia-70m-C-Language-KnowledgeExtract'
device = 'cuda'

model = GPTNeoXForCausalLM.from_pretrained(model_name_or_path).to(device)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

# instruction template
instruction = '[Summarize the knowledge points in the code below]\n'
# any c-lang pieces you like, could be partial functions or statements
input_content = '''```c
int partition(int arr[], int low, int high) {
    int pivot = arr[high];
    int i = (low - 1);
    for (int j = low; j <= high - 1; j++) {
        if (arr[j] < pivot) {
            i++;
            swap(&arr[i], &arr[j]);
        }
    }
    swap(&arr[i + 1], &arr[high]);
    return (i + 1);
}

void quickSort(int arr[], int low, int high) {
    if (low < high) {
        int pi = partition(arr, low, high);
        quickSort(arr, low, pi - 1);
        quickSort(arr, pi + 1, high);
    }
}
```'''
text = instruction + input_content
```

Then generate:

```python
inputs = tokenizer(text, return_tensors="pt").to(device)
tokens = model.generate(
    **inputs,
    pad_token_id=tokenizer.eos_token_id,
    max_new_tokens=32,
)
# deduplicate inputs
response = tokenizer.decode(tokens[0]).split('```')[-1].split('<')[0]
```



### and more

However, in practical use, in order to achieve more diverse representations, it's recommended to do multiple inferences. Don't worry, it's really small so the inferences don't take much time, as follows:

```python
ans_dict = {}
def increment_insert(key):
    ans_dict[key] = ans_dict.get(key, 0) + 1

for i in range(30):		# maybe 20 times or less enough too
    inputs = tokenizer(text, return_tensors="pt").to(device)
    tokens = model.generate(
        **inputs,
        pad_token_id=tokenizer.eos_token_id,
        max_new_tokens=32,
        do_sample=True,
        temperature=2.0,  # high temperature for diversity
        top_p=0.95,
        top_k=30,
    )
    response = tokenizer.decode(tokens[0]).split('```')[-1].split('<')[0]
    increment_insert(response)

print(ans_dict)
### output as below, could take high-freq answers
### {
###     'Backtracking': 1,
###     'Heap': 1,
###     'Quick sort': 25,
###     'Recurrence': 2,
###     'Queue': 1
### }
```