Update README.md
Browse files
README.md
CHANGED
@@ -44,22 +44,23 @@ During the code generation step, we use early stopping on the `}\n}` sequence to
|
|
44 |
The early stopping method, post-processing steps, and evaluation code are available in the example below.
|
45 |
|
46 |
```python
|
47 |
-
import
|
48 |
-
import jsonlines
|
49 |
import re
|
50 |
-
|
|
|
|
|
|
|
51 |
from transformers import (
|
52 |
AutoTokenizer,
|
53 |
AutoModelForCausalLM,
|
54 |
StoppingCriteria,
|
55 |
StoppingCriteriaList,
|
56 |
)
|
57 |
-
from
|
58 |
from mxeval.evaluation import evaluate_functional_correctness
|
59 |
-
|
60 |
|
61 |
class StoppingCriteriaSub(StoppingCriteria):
|
62 |
-
|
63 |
def __init__(self, stops, tokenizer):
|
64 |
(StoppingCriteria.__init__(self),)
|
65 |
self.stops = rf"{stops}"
|
@@ -75,16 +76,8 @@ class StoppingCriteriaSub(StoppingCriteria):
|
|
75 |
|
76 |
|
77 |
def generate(problem):
|
78 |
-
|
79 |
-
|
80 |
-
stopping_criteria = StoppingCriteriaList(
|
81 |
-
[
|
82 |
-
StoppingCriteriaSub(
|
83 |
-
stops= "\n}\n", tokenizer=tokenizer
|
84 |
-
)
|
85 |
-
]
|
86 |
-
)
|
87 |
-
|
88 |
|
89 |
problem = tokenizer.encode(problem, return_tensors="pt").to('cuda')
|
90 |
sample = model.generate(
|
@@ -99,22 +92,20 @@ def generate(problem):
|
|
99 |
)
|
100 |
|
101 |
answer = tokenizer.decode(sample[0], skip_special_tokens=True)
|
102 |
-
|
103 |
return answer
|
104 |
|
105 |
-
def clean_asnwer(code):
|
106 |
|
|
|
107 |
# Clean comments
|
108 |
code_without_line_comments = re.sub(r"//.*", "", code)
|
109 |
code_without_all_comments = re.sub(
|
110 |
r"/\*.*?\*/", "", code_without_line_comments, flags=re.DOTALL
|
111 |
)
|
112 |
-
|
113 |
#Clean signatures
|
114 |
lines = code.split("\n")
|
115 |
for i, line in enumerate(lines):
|
116 |
if line.startswith("fun "):
|
117 |
-
return "\n".join(lines[i + 1
|
118 |
|
119 |
return code
|
120 |
|
@@ -123,10 +114,9 @@ model_name = "JetBrains/CodeLlama-7B-Kexer"
|
|
123 |
dataset = load_dataset("jetbrains/Kotlin_HumanEval")['train']
|
124 |
problem_dict = {problem['task_id']: problem for problem in dataset}
|
125 |
|
126 |
-
model = AutoModelForCausalLM.from_pretrained(model_name,torch_dtype=torch.bfloat16).to('cuda')
|
127 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
128 |
|
129 |
-
|
130 |
output = []
|
131 |
for key in tqdm(list(problem_dict.keys()), leave=False):
|
132 |
problem = problem_dict[key]["prompt"]
|
@@ -134,7 +124,6 @@ for key in tqdm(list(problem_dict.keys()), leave=False):
|
|
134 |
answer = clean_asnwer(answer)
|
135 |
output.append({"task_id": key, "completion": answer, "language": "kotlin"})
|
136 |
|
137 |
-
|
138 |
output_file = f"answers"
|
139 |
with jsonlines.open(output_file, mode="w") as writer:
|
140 |
for line in output:
|
@@ -148,6 +137,17 @@ evaluate_functional_correctness(
|
|
148 |
problem_file=problem_dict,
|
149 |
)
|
150 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
151 |
```
|
152 |
|
153 |
|
|
|
44 |
The early stopping method, post-processing steps, and evaluation code are available in the example below.
|
45 |
|
46 |
```python
|
47 |
+
import json
|
|
|
48 |
import re
|
49 |
+
|
50 |
+
from datasets import load_dataset
|
51 |
+
import jsonlines
|
52 |
+
import torch
|
53 |
from transformers import (
|
54 |
AutoTokenizer,
|
55 |
AutoModelForCausalLM,
|
56 |
StoppingCriteria,
|
57 |
StoppingCriteriaList,
|
58 |
)
|
59 |
+
from tqdm import tqdm
|
60 |
from mxeval.evaluation import evaluate_functional_correctness
|
61 |
+
|
62 |
|
63 |
class StoppingCriteriaSub(StoppingCriteria):
|
|
|
64 |
def __init__(self, stops, tokenizer):
|
65 |
(StoppingCriteria.__init__(self),)
|
66 |
self.stops = rf"{stops}"
|
|
|
76 |
|
77 |
|
78 |
def generate(problem):
|
79 |
+
criterion = StoppingCriteriaSub(stops="\n}\n", tokenizer=tokenizer)
|
80 |
+
stopping_criteria = StoppingCriteriaList([criterion])
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
|
82 |
problem = tokenizer.encode(problem, return_tensors="pt").to('cuda')
|
83 |
sample = model.generate(
|
|
|
92 |
)
|
93 |
|
94 |
answer = tokenizer.decode(sample[0], skip_special_tokens=True)
|
|
|
95 |
return answer
|
96 |
|
|
|
97 |
|
98 |
+
def clean_asnwer(code):
|
99 |
# Clean comments
|
100 |
code_without_line_comments = re.sub(r"//.*", "", code)
|
101 |
code_without_all_comments = re.sub(
|
102 |
r"/\*.*?\*/", "", code_without_line_comments, flags=re.DOTALL
|
103 |
)
|
|
|
104 |
#Clean signatures
|
105 |
lines = code.split("\n")
|
106 |
for i, line in enumerate(lines):
|
107 |
if line.startswith("fun "):
|
108 |
+
return "\n".join(lines[i + 1:])
|
109 |
|
110 |
return code
|
111 |
|
|
|
114 |
dataset = load_dataset("jetbrains/Kotlin_HumanEval")['train']
|
115 |
problem_dict = {problem['task_id']: problem for problem in dataset}
|
116 |
|
117 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to('cuda')
|
118 |
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
119 |
|
|
|
120 |
output = []
|
121 |
for key in tqdm(list(problem_dict.keys()), leave=False):
|
122 |
problem = problem_dict[key]["prompt"]
|
|
|
124 |
answer = clean_asnwer(answer)
|
125 |
output.append({"task_id": key, "completion": answer, "language": "kotlin"})
|
126 |
|
|
|
127 |
output_file = f"answers"
|
128 |
with jsonlines.open(output_file, mode="w") as writer:
|
129 |
for line in output:
|
|
|
137 |
problem_file=problem_dict,
|
138 |
)
|
139 |
|
140 |
+
with open(output_file + '_results.jsonl') as fp:
|
141 |
+
total = 0
|
142 |
+
correct = 0
|
143 |
+
for line in fp:
|
144 |
+
sample_res = json.loads(line)
|
145 |
+
print(sample_res)
|
146 |
+
total += 1
|
147 |
+
correct += sample_res['passed']
|
148 |
+
|
149 |
+
print(f'Pass rate: {correct/total}')
|
150 |
+
|
151 |
```
|
152 |
|
153 |
|