JetBrains/Kotlin_HumanEval

Benchmark summary

We introduce HumanEval for Kotlin, created from scratch by human experts. 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. The tests we implement are equivalent to the original HumanEval tests for Python.

How to use

The benchmark is prepared in a format suitable for MXEval and can be easily integrated into the MXEval pipeline.

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.

The code for running an example model on the benchmark using the early stopping and post-processing is available below.

import json
import re

from datasets import load_dataset
import jsonlines
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    StoppingCriteria,
    StoppingCriteriaList,
)
from tqdm import tqdm 
from mxeval.evaluation import evaluate_functional_correctness


class StoppingCriteriaSub(StoppingCriteria):
    def __init__(self, stops, tokenizer):
        (StoppingCriteria.__init__(self),)
        self.stops = rf"{stops}"
        self.tokenizer = tokenizer

    def __call__(
        self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs
    ) -> bool:
        last_three_tokens = [int(x) for x in input_ids.data[0][-3:]]
        decoded_last_three_tokens = self.tokenizer.decode(last_three_tokens)

        return bool(re.search(self.stops, decoded_last_three_tokens))


def generate(problem):
    criterion = StoppingCriteriaSub(stops="\n}\n", tokenizer=tokenizer)
    stopping_criteria = StoppingCriteriaList([criterion])
    
    problem = tokenizer.encode(problem, return_tensors="pt").to('cuda')
    sample = model.generate(
        problem,
        max_new_tokens=256,
        min_new_tokens=128,
        pad_token_id=tokenizer.eos_token_id,
        do_sample=False,
        num_beams=1,
        stopping_criteria=stopping_criteria,
    )
    
    answer = tokenizer.decode(sample[0], skip_special_tokens=True)
    return answer


def clean_asnwer(code):
    # Clean comments
    code_without_line_comments = re.sub(r"//.*", "", code)
    code_without_all_comments = re.sub(
        r"/\*.*?\*/", "", code_without_line_comments, flags=re.DOTALL
    )
    #Clean signatures
    lines = code.split("\n")
    for i, line in enumerate(lines):
        if line.startswith("fun "):
            return "\n".join(lines[i + 1:])
            
    return code


model_name = "JetBrains/CodeLlama-7B-Kexer"
dataset = load_dataset("jetbrains/Kotlin_HumanEval")['train']
problem_dict = {problem['task_id']: problem for problem in dataset}

model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16).to('cuda')
tokenizer = AutoTokenizer.from_pretrained(model_name)

output = []
for key in tqdm(list(problem_dict.keys()), leave=False):
    problem = problem_dict[key]["prompt"]
    answer = generate(problem)
    answer = clean_asnwer(answer)
    output.append({"task_id": key, "completion": answer, "language": "kotlin"})

output_file = f"answers"
with jsonlines.open(output_file, mode="w") as writer:
    for line in output:
        writer.write(line)

evaluate_functional_correctness(
    sample_file=output_file,
    k=[1],
    n_workers=16,
    timeout=15,
    problem_file=problem_dict,
)

with open(output_file + '_results.jsonl') as fp:
    total = 0
    correct = 0
    for line in fp:
        sample_res = json.loads(line)
        print(sample_res)
        total += 1
        correct += sample_res['passed']

print(f'Pass rate: {correct/total}')

Results

We evaluated multiple coding models using this benchmark, and the results are presented in the figure below: