Commit
·
8b6caa7
0
Parent(s):
Init commit
Browse files- README.md +156 -0
- data/train-00000-of-00001.parquet +3 -0
- model_scores.png +3 -0
README.md
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
dataset_info:
|
4 |
+
features:
|
5 |
+
- name: task_id
|
6 |
+
dtype: string
|
7 |
+
- name: prompt
|
8 |
+
dtype: string
|
9 |
+
- name: entry_point
|
10 |
+
dtype: string
|
11 |
+
- name: test
|
12 |
+
dtype: string
|
13 |
+
- name: description
|
14 |
+
dtype: string
|
15 |
+
- name: language
|
16 |
+
dtype: string
|
17 |
+
- name: canonical_solution
|
18 |
+
sequence: string
|
19 |
+
splits:
|
20 |
+
- name: train
|
21 |
+
num_bytes: 505355
|
22 |
+
num_examples: 161
|
23 |
+
download_size: 174830
|
24 |
+
dataset_size: 505355
|
25 |
+
configs:
|
26 |
+
- config_name: default
|
27 |
+
data_files:
|
28 |
+
- split: train
|
29 |
+
path: data/train-*
|
30 |
+
---
|
31 |
+
|
32 |
+
# Benchmark summary
|
33 |
+
|
34 |
+
We introduce HumanEval for Kotlin, created from scratch by human experts.
|
35 |
+
Solutions and tests for all 161 HumanEval tasks are written by an expert olympiad programmer with 6 years of experience in Kotlin, and independently checked by a programmer with 4 years of experience in Kotlin.
|
36 |
+
The tests we implement are eqivalent to the original HumanEval tests for Python.
|
37 |
+
|
38 |
+
# How to use
|
39 |
+
|
40 |
+
The benchmark is prepared in a format suitable for MXEval and can be easily integrated into the MXEval pipeline.
|
41 |
+
|
42 |
+
When testing models on this benchmark, during the code generation step we use early stopping on the `}\n}` sequence to expedite the process. We also perform some code post-processing before evaluation — specifically, we remove all comments and signatures.
|
43 |
+
|
44 |
+
The code for running an example model on the benchmark using the early stopping and post-processing is available 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}"
|
67 |
+
self.tokenizer = tokenizer
|
68 |
+
|
69 |
+
def __call__(
|
70 |
+
self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
|
71 |
+
) -> bool:
|
72 |
+
last_three_tokens = [int(x) for x in input_ids.data[0][-3:]]
|
73 |
+
decoded_last_three_tokens = self.tokenizer.decode(last_three_tokens)
|
74 |
+
|
75 |
+
return bool(re.search(self.stops, decoded_last_three_tokens))
|
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(
|
84 |
+
problem,
|
85 |
+
max_new_tokens=256,
|
86 |
+
min_new_tokens=128,
|
87 |
+
pad_token_id=tokenizer.eos_token_id,
|
88 |
+
do_sample=False,
|
89 |
+
num_beams=1,
|
90 |
+
stopping_criteria=stopping_criteria,
|
91 |
+
)
|
92 |
+
|
93 |
+
answer = tokenizer.decode(sample[0], skip_special_tokens=True)
|
94 |
+
return answer
|
95 |
+
|
96 |
+
|
97 |
+
def clean_asnwer(code):
|
98 |
+
# Clean comments
|
99 |
+
code_without_line_comments = re.sub(r"//.*", "", code)
|
100 |
+
code_without_all_comments = re.sub(
|
101 |
+
r"/\*.*?\*/", "", code_without_line_comments, flags=re.DOTALL
|
102 |
+
)
|
103 |
+
#Clean signatures
|
104 |
+
lines = code.split("\n")
|
105 |
+
for i, line in enumerate(lines):
|
106 |
+
if line.startswith("fun "):
|
107 |
+
return "\n".join(lines[i + 1:])
|
108 |
+
|
109 |
+
return code
|
110 |
+
|
111 |
+
|
112 |
+
model_name = "JetBrains/CodeLlama-7B-Kexer"
|
113 |
+
dataset = load_dataset("jetbrains/Kotlin_HumanEval")['train']
|
114 |
+
problem_dict = {problem['task_id']: problem for problem in dataset}
|
115 |
+
|
116 |
+
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to('cuda')
|
117 |
+
tokenizer = AutoTokenizer.from_pretrained(model_name)
|
118 |
+
|
119 |
+
output = []
|
120 |
+
for key in tqdm(list(problem_dict.keys()), leave=False):
|
121 |
+
problem = problem_dict[key]["prompt"]
|
122 |
+
answer = generate(problem)
|
123 |
+
answer = clean_asnwer(answer)
|
124 |
+
output.append({"task_id": key, "completion": answer, "language": "kotlin"})
|
125 |
+
|
126 |
+
output_file = f"answers"
|
127 |
+
with jsonlines.open(output_file, mode="w") as writer:
|
128 |
+
for line in output:
|
129 |
+
writer.write(line)
|
130 |
+
|
131 |
+
evaluate_functional_correctness(
|
132 |
+
sample_file=output_file,
|
133 |
+
k=[1],
|
134 |
+
n_workers=16,
|
135 |
+
timeout=15,
|
136 |
+
problem_file=problem_dict,
|
137 |
+
)
|
138 |
+
|
139 |
+
with open(output_file + '_results.jsonl') as fp:
|
140 |
+
total = 0
|
141 |
+
correct = 0
|
142 |
+
for line in fp:
|
143 |
+
sample_res = json.loads(line)
|
144 |
+
print(sample_res)
|
145 |
+
total += 1
|
146 |
+
correct += sample_res['passed']
|
147 |
+
|
148 |
+
print(f'Pass rate: {correct/total}')
|
149 |
+
|
150 |
+
```
|
151 |
+
|
152 |
+
|
153 |
+
# Results
|
154 |
+
|
155 |
+
We evaluated multiple coding models using this benchmark, and the results are presented in the figure below:
|
156 |
+

|
data/train-00000-of-00001.parquet
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:54c2765720acf0887c603f1eb1c08382cb7eba0c39a7a37be8e3b982aaddbc7b
|
3 |
+
size 174830
|
model_scores.png
ADDED
![]() |
Git LFS Details
|