Datasets:

Modalities:
Text
Formats:
parquet
Size:
< 1K
Libraries:
Datasets
pandas
License:
parquet-converter commited on
Commit
cc43ca5
1 Parent(s): 8b6caa7

Update parquet files

Browse files
.gitattributes ADDED
@@ -0,0 +1 @@
 
 
1
+ default/train/0000.parquet filter=lfs diff=lfs merge=lfs -text
README.md DELETED
@@ -1,156 +0,0 @@
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
- ![results](https://huggingface.co/datasets/JetBrains/Kotlin_HumanEval/resolve/main/model_scores.png)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
data/train-00000-of-00001.parquet → default/train/0000.parquet RENAMED
File without changes
model_scores.png DELETED

Git LFS Details

  • SHA256: 7c6bf5e9b7863dd5e21683f8773c8beb3a74e4c1c46fdb55c46de18b8e0e9c65
  • Pointer size: 131 Bytes
  • Size of remote file: 441 kB