{"task": "code_en_03952e9e75dec136b2594301", "id": "03952e9e75dec136b2594301", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/43", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef pairs_sum_to_zero(l):\n \"\"\"\n pairs_sum_to_zero takes a list of integers as an input.\n it returns True if there are two distinct elements in the list that\n sum to zero, and False otherwise.\n >>> pairs_sum_to_zero([1, 3, 5, 0])\n False\n >>> pairs_sum_to_zero([1, 3, -2, 1])\n False\n >>> pairs_sum_to_zero([1, 2, 3, 7])\n False\n >>> pairs_sum_to_zero([2, 4, -5, 3, 5, 7])\n True\n >>> pairs_sum_to_zero([1])\n False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core semantics: returns True exactly when two elements at distinct indices sum to zero (e.g., [2,4,-5,3,5,7] -> True via -5 and 5), and False for all docstring False examples, including [1,3,-2,1] where no x/-x pair exists. A single 0 must not pair with itself ([1,3,5,0] -> False), but two zeros at different positions ([0,0]) should return True. Wrong results on docstring examples are severe; wrong handling of the duplicate-zero case is a moderate deduction."}, {"id": "c2", "description": "Boundary inputs: handles the empty list and single-element lists ([1], [0]) by returning False without crashing (no IndexError or similar). Handles negative numbers, duplicates, and large integer magnitudes correctly. Any crash or non-termination on a valid list of integers scores 6 or below."}, {"id": "c3", "description": "Algorithmic suitability: uses an approach whose results are correct, such as a nested index loop with i>> pairs_sum_to_zero([1, 3, 5, 0])\n False\n >>> pairs_sum_to_zero([1, 3, -2, 1])\n False\n >>> pairs_sum_to_zero([1, 2, 3, 7])\n False\n >>> pairs_sum_to_zero([2, 4, -5, 3, 5, 7])\n True\n >>> pairs_sum_to_zero([1])\n False\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([1, 3, 5, 0]) == False\n assert candidate([1, 3, -2, 1]) == False\n assert candidate([1, 2, 3, 7]) == False\n assert candidate([2, 4, -5, 3, 5, 7]) == True\n assert candidate([1]) == False\n\n assert candidate([-3, 9, -1, 3, 2, 30]) == True\n assert candidate([-3, 9, -1, 3, 2, 31]) == True\n assert candidate([-3, 9, -1, 4, 2, 30]) == False\n assert candidate([-3, 9, -1, 4, 2, 31]) == False\n\n", "entry_point": "pairs_sum_to_zero"} {"task": "code_en_062346c3ba709a11c6d3d7de", "id": "062346c3ba709a11c6d3d7de", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/14", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\nfrom typing import List\n\n\ndef all_prefixes(string: str) -> List[str]:\n \"\"\" Return list of all prefixes from shortest to longest of the input string\n >>> all_prefixes('abc')\n ['a', 'ab', 'abc']\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: For a non-empty string, the body returns a list of exactly len(string) prefixes string[:1] through string[:n], ordered shortest to longest. For 'abc' it returns ['a','ab','abc']. The empty prefix '' is excluded."}, {"id": "c2", "description": "Boundary cases: An empty input string returns an empty list [] without crashing. A single-character string returns a one-element list. Strings with repeated characters, spaces or unicode keep every prefix and do not deduplicate any. Long strings terminate, and the expected quadratic total output size is acceptable."}, {"id": "c3", "description": "Validity: The returned text is only the indented function body, with no repeated signature, docstring or Markdown fences. The first line's indentation is consistent with the rest, so the body is syntactically valid when placed under the given signature. The function returns a List[str] and does not print."}, {"id": "c4", "description": "Algorithm and readability: The approach is a simple, direct method such as slicing in a loop or a list comprehension over range(1, len(string)+1). It has no unnecessary complexity, no extra guards that change outputs, and no added requirements. Any correct alternative algorithm receives full credit."}]}, "target_words": null, "reference": " result = []\n\n for i in range(len(string)):\n result.append(string[:i+1])\n return result\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/14"}, "prefix": "from typing import List\n\n\ndef all_prefixes(string: str) -> List[str]:\n \"\"\" Return list of all prefixes from shortest to longest of the input string\n >>> all_prefixes('abc')\n ['a', 'ab', 'abc']\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('') == []\n assert candidate('asdfgh') == ['a', 'as', 'asd', 'asdf', 'asdfg', 'asdfgh']\n assert candidate('WWW') == ['W', 'WW', 'WWW']\n", "entry_point": "all_prefixes"} {"task": "code_en_109a97831230e7ba6e2c37be", "id": "109a97831230e7ba6e2c37be", "domain": "code", "language": "en", "split": "validation", "group": "HumanEval/161", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef solve(s):\n \"\"\"You are given a string s.\n if s[i] is a letter, reverse its case from lower to upper or vise versa, \n otherwise keep it as it is.\n If the string contains no letters, reverse the string.\n The function should return the resulted string.\n Examples\n solve(\"1234\") = \"4321\"\n solve(\"ab\") = \"AB\"\n solve(\"#a@C\") = \"#A@c\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Case swapping: when s contains at least one letter, every letter has its case inverted and every non-letter character stays unchanged and in its original position, with no reversal. Mentally check that solve(\"ab\") returns \"AB\" and solve(\"#a@C\") returns \"#A@c\"."}, {"id": "c2", "description": "No-letter reversal: when s contains no letters, the whole string is returned reversed. Check that solve(\"1234\") returns \"4321\". The empty string must return \"\" without crashing. Single-character and all-symbol strings such as \"#@!\" must be handled correctly."}, {"id": "c3", "description": "Letter detection: letters are detected with str.isalpha() or an equivalent. Swapping uses swapcase(), or per-character upper()/lower() applied only to letters. The letter/no-letter decision covers the whole string and is not made per character. The solution runs in linear time."}, {"id": "c4", "description": "The returned text is a valid, correctly indented function body. It has no signature, docstring or Markdown fences, and it returns a string in every path. Readability is a minor factor. Deduct heavily for syntax or indentation errors."}]}, "target_words": null, "reference": " flg = 0\n idx = 0\n new_str = list(s)\n for i in s:\n if i.isalpha():\n new_str[idx] = i.swapcase()\n flg = 1\n idx += 1\n s = \"\"\n for i in new_str:\n s += i\n if flg == 0:\n return s[len(s)::-1]\n return s\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/161"}, "prefix": "\ndef solve(s):\n \"\"\"You are given a string s.\n if s[i] is a letter, reverse its case from lower to upper or vise versa, \n otherwise keep it as it is.\n If the string contains no letters, reverse the string.\n The function should return the resulted string.\n Examples\n solve(\"1234\") = \"4321\"\n solve(\"ab\") = \"AB\"\n solve(\"#a@C\") = \"#A@c\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"AsDf\") == \"aSdF\"\n assert candidate(\"1234\") == \"4321\"\n assert candidate(\"ab\") == \"AB\"\n assert candidate(\"#a@C\") == \"#A@c\"\n assert candidate(\"#AsdfW^45\") == \"#aSDFw^45\"\n assert candidate(\"#6@2\") == \"2@6#\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(\"#$a^D\") == \"#$A^d\"\n assert candidate(\"#ccc\") == \"#CCC\"\n\n # Don't remove this line:\n", "entry_point": "solve"} {"task": "code_en_11724d4a3dd81a33444c8377", "id": "11724d4a3dd81a33444c8377", "domain": "code", "language": "en", "split": "validation", "group": "HumanEval/52", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef below_threshold(l: list, t: int):\n \"\"\"Return True if all numbers in the list l are below threshold t.\n >>> below_threshold([1, 2, 4, 10], 100)\n True\n >>> below_threshold([1, 20, 4, 10], 5)\n False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Functional correctness: the body returns True exactly when every element of l is strictly less than t, and False otherwise. Elements equal to t must yield False, so below_threshold([5], 5) is False. Both docstring examples must produce the shown results."}, {"id": "c2", "description": "Boundary cases: an empty list returns True, which is vacuous truth. Negative numbers, zero, floats mixed with ints, and a single-element list are all handled correctly. The body terminates and does not crash on any list of numbers. No guards are added that change outputs, such as raising errors on an empty list."}, {"id": "c3", "description": "Validity and syntax: the returned text is only an indented function body, with no repeated signature or docstring and no Markdown. Its indentation is consistent, starting at the function-body level, so it parses and runs when appended under the given signature."}, {"id": "c4", "description": "Algorithmic suitability and readability: the body uses a linear single pass, such as all(x < t for x in l) or a loop with early return. It does not use needless sorting, copying, or extra complexity. The code is clear and concise. Different correct approaches receive full credit."}]}, "target_words": null, "reference": " for e in l:\n if e >= t:\n return False\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/52"}, "prefix": "\n\ndef below_threshold(l: list, t: int):\n \"\"\"Return True if all numbers in the list l are below threshold t.\n >>> below_threshold([1, 2, 4, 10], 100)\n True\n >>> below_threshold([1, 20, 4, 10], 5)\n False\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([1, 2, 4, 10], 100)\n assert not candidate([1, 20, 4, 10], 5)\n assert candidate([1, 20, 4, 10], 21)\n assert candidate([1, 20, 4, 10], 22)\n assert candidate([1, 8, 4, 10], 11)\n assert not candidate([1, 8, 4, 10], 10)\n\n", "entry_point": "below_threshold"} {"task": "code_en_15bc8a32dfa3f020330f1819", "id": "15bc8a32dfa3f020330f1819", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/75", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef is_multiply_prime(a):\n \"\"\"Write a function that returns true if the given number is the multiplication of 3 prime numbers\n and false otherwise.\n Knowing that (a) is less then 100. \n Example:\n is_multiply_prime(30) == True\n 30 = 2 * 3 * 5\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: returns True exactly when a equals a product of three primes (repetition allowed, e.g. 8=2*2*2, 12=2*2*3, 30=2*3*5) and False otherwise (e.g. primes, 1, 0, 4, 6, 16, 36). It must produce correct values for all a < 100. Wrong results on common inputs like 8 or 30 warrant a low score. A different correct algorithm, such as factor counting or a triple loop over primes, receives full credit."}, {"id": "c2", "description": "Boundary robustness: handles a = 0, 1, 2 and negative values without crashing or looping forever. It returns False for these inputs. A crash or non-termination on any integer below 100 scores 6 or below."}, {"id": "c3", "description": "Validity: the output is only an indented function body with consistent indentation and valid syntax. It has no repeated signature or Markdown, and it uses a correct, terminating primality test or factorization."}, {"id": "c4", "description": "Readability and suitability: the logic is clear and appropriately efficient for small inputs. It has no extraneous guards that change outputs and no needless complexity."}]}, "target_words": null, "reference": " def is_prime(n):\n for j in range(2,n):\n if n%j == 0:\n return False\n return True\n\n for i in range(2,101):\n if not is_prime(i): continue\n for j in range(2,101):\n if not is_prime(j): continue\n for k in range(2,101):\n if not is_prime(k): continue\n if i*j*k == a: return True\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/75"}, "prefix": "\ndef is_multiply_prime(a):\n \"\"\"Write a function that returns true if the given number is the multiplication of 3 prime numbers\n and false otherwise.\n Knowing that (a) is less then 100. \n Example:\n is_multiply_prime(30) == True\n 30 = 2 * 3 * 5\n \"\"\"\n", "tests": "def check(candidate):\n\n assert candidate(5) == False\n assert candidate(30) == True\n assert candidate(8) == True\n assert candidate(10) == False\n assert candidate(125) == True\n assert candidate(3 * 5 * 7) == True\n assert candidate(3 * 6 * 7) == False\n assert candidate(9 * 9 * 9) == False\n assert candidate(11 * 9 * 9) == False\n assert candidate(11 * 13 * 7) == True\n\n", "entry_point": "is_multiply_prime"} {"task": "code_en_1e8c7fcdbd98de37d71cbc16", "id": "1e8c7fcdbd98de37d71cbc16", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/99", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef closest_integer(value):\n '''\n Create a function that takes a value (string) representing a number\n and returns the closest integer to it. If the number is equidistant\n from two integers, round it away from zero.\n\n Examples\n >>> closest_integer(\"10\")\n 10\n >>> closest_integer(\"15.3\")\n 15\n\n Note:\n Rounding away from zero means that if the given number is equidistant\n from two integers, the one you should return is the one that is the\n farthest from zero. For example closest_integer(\"14.5\") should\n return 15 and closest_integer(\"-14.5\") should return -15.\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns the correct nearest integer as an int for ordinary inputs, e.g. \"10\"->10, \"15.3\"->15, \"15.7\"->16, \"-15.3\"->-15, \"-15.7\"->-16. It must not rely on Python's banker's round(), which gives round(14.5)=14."}, {"id": "c2", "description": "Handles exact halfway cases by rounding away from zero: \"14.5\"->15, \"-14.5\"->-15, \"0.5\"->1, \"-0.5\"->-1, \"2.5\"->3. Also handles inputs with trailing zeros such as \"14.50\" and inputs like \"0\" and \"-0\"."}, {"id": "c3", "description": "Is robust on input forms within the domain without crashing. This covers integer strings without a decimal point, negative numbers, and an empty string (if the code handles it, it should return 0 rather than crash). It also covers large values, where float precision issues near .5 (e.g. long decimal strings) are a moderate concern, and should not crash."}, {"id": "c4", "description": "The body is syntactically valid and correctly indented as a function body, with no signature or Markdown. It uses a clear, appropriate approach such as floor(x+0.5)/ceil(x-0.5) by sign or Decimal ROUND_HALF_UP, and avoids unnecessary complexity."}]}, "target_words": null, "reference": " from math import floor, ceil\n\n if value.count('.') == 1:\n # remove trailing zeros\n while (value[-1] == '0'):\n value = value[:-1]\n\n num = float(value)\n if value[-2:] == '.5':\n if num > 0:\n res = ceil(num)\n else:\n res = floor(num)\n elif len(value) > 0:\n res = int(round(num))\n else:\n res = 0\n\n return res\n\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/99"}, "prefix": "\ndef closest_integer(value):\n '''\n Create a function that takes a value (string) representing a number\n and returns the closest integer to it. If the number is equidistant\n from two integers, round it away from zero.\n\n Examples\n >>> closest_integer(\"10\")\n 10\n >>> closest_integer(\"15.3\")\n 15\n\n Note:\n Rounding away from zero means that if the given number is equidistant\n from two integers, the one you should return is the one that is the\n farthest from zero. For example closest_integer(\"14.5\") should\n return 15 and closest_integer(\"-14.5\") should return -15.\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"10\") == 10, \"Test 1\"\n assert candidate(\"14.5\") == 15, \"Test 2\"\n assert candidate(\"-15.5\") == -16, \"Test 3\"\n assert candidate(\"15.3\") == 15, \"Test 3\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(\"0\") == 0, \"Test 0\"\n\n", "entry_point": "closest_integer"} {"task": "code_en_2212058a2edeaa1b13166f31", "id": "2212058a2edeaa1b13166f31", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/13", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef greatest_common_divisor(a: int, b: int) -> int:\n \"\"\" Return a greatest common divisor of two integers a and b\n >>> greatest_common_divisor(3, 5)\n 1\n >>> greatest_common_divisor(25, 15)\n 5\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields greatest_common_divisor(3, 5) == 1 and greatest_common_divisor(25, 15) == 5, and the function returns the greatest common divisor for general positive integer pairs, including when a < b, a == b, or one argument divides the other. A wrong value on these common inputs should score low."}, {"id": "c2", "description": "Boundary handling: the body terminates without crashing when one argument is 0, e.g. gcd(0, 5) == 5 and gcd(5, 0) == 5, and handles gcd(0, 0) without an infinite loop or ZeroDivisionError. Returning 0 is the conventional result for gcd(0, 0). For negative inputs, it terminates and returns a value whose magnitude is the true gcd. Non-negative output is preferred, but a sign difference on negatives is only a moderate issue because the docstring is ambiguous. Any non-termination or crash on these integer inputs scores 6 or below."}, {"id": "c3", "description": "Algorithmic suitability: the body uses an efficient method, such as the Euclidean algorithm (modulo or iterative), math.gcd, or an equivalent. It handles large values, e.g. around 10**18, quickly. Brute-force trial division down from min(a, b) is acceptable only if it is correct, and it should be penalized for impractical runtime on large inputs. Any correct algorithm gets full credit."}, {"id": "c4", "description": "Validity and readability: the output is only an indented function body with no signature, docstring repetition, or Markdown fences. Indentation is consistent and syntactically valid when placed under the given signature, and if the first line's indentation breaks the body, this criterion is lowered. Any imports are placed inside the body. The code is concise and clear. Extra commentary is not rewarded."}]}, "target_words": null, "reference": " while b:\n a, b = b, a % b\n return a\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/13"}, "prefix": "\n\ndef greatest_common_divisor(a: int, b: int) -> int:\n \"\"\" Return a greatest common divisor of two integers a and b\n >>> greatest_common_divisor(3, 5)\n 1\n >>> greatest_common_divisor(25, 15)\n 5\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate(3, 7) == 1\n assert candidate(10, 15) == 5\n assert candidate(49, 14) == 7\n assert candidate(144, 60) == 12\n", "entry_point": "greatest_common_divisor"} {"task": "code_en_22ab90c8fd12d95583556c51", "id": "22ab90c8fd12d95583556c51", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/109", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef move_one_ball(arr):\n \"\"\"We have an array 'arr' of N integers arr[1], arr[2], ..., arr[N].The\n numbers in the array will be randomly ordered. Your task is to determine if\n it is possible to get an array sorted in non-decreasing order by performing \n the following operation on the given array:\n You are allowed to perform right shift operation any number of times.\n \n One right shift operation means shifting all elements of the array by one\n position in the right direction. The last element of the array will be moved to\n the starting position in the array i.e. 0th index. \n\n If it is possible to obtain the sorted array by performing the above operation\n then return True else return False.\n If the given array is empty then return True.\n\n Note: The given list is guaranteed to have unique elements.\n\n For Example:\n \n move_one_ball([3, 4, 5, 1, 2])==>True\n Explanation: By performin 2 right shift operations, non-decreasing order can\n be achieved for the given array.\n move_one_ball([3, 5, 4, 1, 2])==>False\n Explanation:It is not possible to get non-decreasing order for the given\n array by performing any number of right shift operations.\n \n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly returns True exactly when some rotation of arr is sorted in non-decreasing order, e.g. True for [3,4,5,1,2] and False for [3,5,4,1,2]. Valid approaches include counting descents cyclically (at most one, including the wrap from last to first) or comparing rotations against sorted(arr)."}, {"id": "c2", "description": "Handles boundary inputs without crashing: the empty list returns True, a single-element list returns True, an already sorted list returns True, and a two-element list returns True. Negative numbers are handled correctly, and a fully reversed list of length 3 or more returns False. Any IndexError or non-termination on these inputs is severe."}, {"id": "c3", "description": "The returned body is syntactically valid and consistently indented as a function body, with no repeated signature or Markdown fences. The algorithm is reasonable: O(n) or O(n^2) rotation checking is acceptable. Extra guards must not change the outputs."}, {"id": "c4", "description": "The code is readable, with clear logic and variable names. This is weighted below correctness, and commentary or added complexity earns no extra credit."}]}, "target_words": null, "reference": " if len(arr)==0:\n return True\n sorted_array=sorted(arr)\n my_arr=[]\n \n min_value=min(arr)\n min_index=arr.index(min_value)\n my_arr=arr[min_index:]+arr[0:min_index]\n for i in range(len(arr)):\n if my_arr[i]!=sorted_array[i]:\n return False\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/109"}, "prefix": "\ndef move_one_ball(arr):\n \"\"\"We have an array 'arr' of N integers arr[1], arr[2], ..., arr[N].The\n numbers in the array will be randomly ordered. Your task is to determine if\n it is possible to get an array sorted in non-decreasing order by performing \n the following operation on the given array:\n You are allowed to perform right shift operation any number of times.\n \n One right shift operation means shifting all elements of the array by one\n position in the right direction. The last element of the array will be moved to\n the starting position in the array i.e. 0th index. \n\n If it is possible to obtain the sorted array by performing the above operation\n then return True else return False.\n If the given array is empty then return True.\n\n Note: The given list is guaranteed to have unique elements.\n\n For Example:\n \n move_one_ball([3, 4, 5, 1, 2])==>True\n Explanation: By performin 2 right shift operations, non-decreasing order can\n be achieved for the given array.\n move_one_ball([3, 5, 4, 1, 2])==>False\n Explanation:It is not possible to get non-decreasing order for the given\n array by performing any number of right shift operations.\n \n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([3, 4, 5, 1, 2])==True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([3, 5, 10, 1, 2])==True\n assert candidate([4, 3, 1, 2])==False\n # Check some edge cases that are easy to work out by hand.\n assert candidate([3, 5, 4, 1, 2])==False, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate([])==True\n", "entry_point": "move_one_ball"} {"task": "code_en_2aef681f0d4162c05971d514", "id": "2aef681f0d4162c05971d514", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/16", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef count_distinct_characters(string: str) -> int:\n \"\"\" Given a string, find out how many distinct characters (regardless of case) does it consist of\n >>> count_distinct_characters('xyzXYZ')\n 3\n >>> count_distinct_characters('Jerry')\n 4\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields 3 for 'xyzXYZ' and 4 for 'Jerry', counting characters case-insensitively (e.g. via set of lowercased characters). Any wrong value on these examples scores low."}, {"id": "c2", "description": "Boundary cases: returns 0 for the empty string without crashing; correctly counts non-letter characters (digits, spaces, punctuation) as distinct characters; handles strings of a single repeated character in mixed case (e.g. 'aAaA' -> 1). A crash on any string input scores 6 or below."}, {"id": "c3", "description": "Validity of returned body: output is only the indented function body (no signature, docstring, or Markdown fences), with consistent indentation and valid syntax so it runs when placed under the given signature; it returns an int rather than printing."}, {"id": "c4", "description": "Algorithmic suitability and readability: uses a linear-time approach (e.g. len(set(string.lower()))) rather than quadratic or convoluted logic; code is concise and clear. Any correct alternative case-folding approach (lower/casefold) earns full credit; no extra guards that alter outputs."}]}, "target_words": null, "reference": " return len(set(string.lower()))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/16"}, "prefix": "\n\ndef count_distinct_characters(string: str) -> int:\n \"\"\" Given a string, find out how many distinct characters (regardless of case) does it consist of\n >>> count_distinct_characters('xyzXYZ')\n 3\n >>> count_distinct_characters('Jerry')\n 4\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('') == 0\n assert candidate('abcde') == 5\n assert candidate('abcde' + 'cade' + 'CADE') == 5\n assert candidate('aaaaAAAAaaaa') == 1\n assert candidate('Jerry jERRY JeRRRY') == 5\n", "entry_point": "count_distinct_characters"} {"task": "code_en_2b6a47943a53ce8e034fd129", "id": "2b6a47943a53ce8e034fd129", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/22", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\nfrom typing import List, Any\n\n\ndef filter_integers(values: List[Any]) -> List[int]:\n \"\"\" Filter given list of any python values only for integers\n >>> filter_integers(['a', 3.14, 5])\n [5]\n >>> filter_integers([1, 2, 3, 'abc', {}, []])\n [1, 2, 3]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: the returned body, when executed, yields [5] for ['a', 3.14, 5] and [1, 2, 3] for [1, 2, 3, 'abc', {}, []], preserving original order and duplicates. Any crash on these inputs scores 6 or below."}, {"id": "c2", "description": "Type discrimination edge cases: floats with integral values (e.g. 3.0), numeric strings like '5', None, nested lists, dicts and tuples are excluded. Ints are identified with isinstance(x, int) or an equivalent check, and values are not converted or coerced. Booleans are treated consistently with isinstance semantics. Excluding bools is a minor deviation, not a severe failure."}, {"id": "c3", "description": "Boundary inputs: an empty list returns [], and negative ints, zero and very large ints are retained unchanged. Nothing crashes on any value type, and the input list is not mutated."}, {"id": "c4", "description": "Validity and readability: the output is only a correctly indented function body with no signature, docstring or Markdown. It is syntactically valid and uses a linear-time, clear approach such as a list comprehension. Lower this criterion if the indentation breaks the body."}]}, "target_words": null, "reference": " return [x for x in values if isinstance(x, int)]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/22"}, "prefix": "from typing import List, Any\n\n\ndef filter_integers(values: List[Any]) -> List[int]:\n \"\"\" Filter given list of any python values only for integers\n >>> filter_integers(['a', 3.14, 5])\n [5]\n >>> filter_integers([1, 2, 3, 'abc', {}, []])\n [1, 2, 3]\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate([]) == []\n assert candidate([4, {}, [], 23.2, 9, 'adasd']) == [4, 9]\n assert candidate([3, 'c', 3, 3, 'a', 'b']) == [3, 3, 3]\n", "entry_point": "filter_integers"} {"task": "code_en_2c2b037a21fb2e804c6284fd", "id": "2c2b037a21fb2e804c6284fd", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/8", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\nfrom typing import List, Tuple\n\n\ndef sum_product(numbers: List[int]) -> Tuple[int, int]:\n \"\"\" For a given list of integers, return a tuple consisting of a sum and a product of all the integers in a list.\n Empty sum should be equal to 0 and empty product should be equal to 1.\n >>> sum_product([])\n (0, 1)\n >>> sum_product([1, 2, 3, 4])\n (10, 24)\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns a tuple (sum, product) of all integers in the list. Mentally executing the code gives (0, 1) for [] and (10, 24) for [1, 2, 3, 4]."}, {"id": "c2", "description": "Handles boundary inputs correctly: a single element, e.g. [5] gives (5, 5); zeros, e.g. [0, 3] gives (3, 0); negatives, e.g. [-2, 3] gives (1, -6); and large values, which rely on Python's arbitrary-precision ints with no overflow or truncation. It must not crash on any list of ints. A crash here, such as an initial product of 0 or indexing numbers[0] on an empty list, is a severe failure."}, {"id": "c3", "description": "The returned body is syntactically valid and consistently indented as a function body, with no repeated signature, docstring or Markdown fences. If the first line's indentation breaks the body, this is noted and the validity score is lowered."}, {"id": "c4", "description": "The code uses a straightforward O(n) approach, such as a single loop, sum(), or math.prod/reduce with an initial value of 1, and is readable. It adds no guards that change outputs and has no unnecessary complexity."}]}, "target_words": null, "reference": " sum_value = 0\n prod_value = 1\n\n for n in numbers:\n sum_value += n\n prod_value *= n\n return sum_value, prod_value\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/8"}, "prefix": "from typing import List, Tuple\n\n\ndef sum_product(numbers: List[int]) -> Tuple[int, int]:\n \"\"\" For a given list of integers, return a tuple consisting of a sum and a product of all the integers in a list.\n Empty sum should be equal to 0 and empty product should be equal to 1.\n >>> sum_product([])\n (0, 1)\n >>> sum_product([1, 2, 3, 4])\n (10, 24)\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate([]) == (0, 1)\n assert candidate([1, 1, 1]) == (3, 1)\n assert candidate([100, 0]) == (100, 0)\n assert candidate([3, 5, 7]) == (3 + 5 + 7, 3 * 5 * 7)\n assert candidate([10]) == (10, 10)\n", "entry_point": "sum_product"} {"task": "code_en_3343e5d27e069beb4930ca3a", "id": "3343e5d27e069beb4930ca3a", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/151", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef double_the_difference(lst):\n '''\n Given a list of numbers, return the sum of squares of the numbers\n in the list that are odd. Ignore numbers that are negative or not integers.\n \n double_the_difference([1, 3, 2, 0]) == 1 + 9 + 0 + 0 = 10\n double_the_difference([-1, -2, 0]) == 0\n double_the_difference([9, -2]) == 81\n double_the_difference([0]) == 0 \n \n If the input list is empty, return 0.\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Functional correctness: the body returns the sum of squares of elements that are positive odd integers. It yields 10 for [1,3,2,0], 0 for [-1,-2,0], 81 for [9,-2], 0 for [0] and 0 for []. Negative odd numbers such as -1 and -3 are excluded, and zero and even numbers contribute nothing."}, {"id": "c2", "description": "Non-integer handling: floats such as 3.0, 1.5 or 5.5 are ignored and not squared. An integrality check like isinstance(x, int) satisfies this, as does an equivalent check that excludes floats even when they have integral values. Mixed lists do not crash. A crash on a float input counts as a severe failure. Treatment of bool values is a rare edge case and gets at most a minor penalty."}, {"id": "c3", "description": "Validity and syntax: the output is only an indented function body with consistent indentation, starting at the correct body level. It has no repeated signature, no Markdown fences and no syntax errors. It parses and runs as the body of the given def."}, {"id": "c4", "description": "Algorithmic suitability and readability: the body uses a single linear pass, for example a sum over a generator with a clear filter condition. It has no unnecessary complexity and no extra guards that change outputs. Any correct alternative algorithm gets full credit."}]}, "target_words": null, "reference": " return sum([i**2 for i in lst if i > 0 and i%2!=0 and \".\" not in str(i)])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/151"}, "prefix": "\ndef double_the_difference(lst):\n '''\n Given a list of numbers, return the sum of squares of the numbers\n in the list that are odd. Ignore numbers that are negative or not integers.\n \n double_the_difference([1, 3, 2, 0]) == 1 + 9 + 0 + 0 = 10\n double_the_difference([-1, -2, 0]) == 0\n double_the_difference([9, -2]) == 81\n double_the_difference([0]) == 0 \n \n If the input list is empty, return 0.\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([]) == 0 , \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([5, 4]) == 25 , \"This prints if this assert fails 2 (good for debugging!)\"\n assert candidate([0.1, 0.2, 0.3]) == 0 , \"This prints if this assert fails 3 (good for debugging!)\"\n assert candidate([-10, -20, -30]) == 0 , \"This prints if this assert fails 4 (good for debugging!)\"\n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([-1, -2, 8]) == 0, \"This prints if this assert fails 5 (also good for debugging!)\"\n assert candidate([0.2, 3, 5]) == 34, \"This prints if this assert fails 6 (also good for debugging!)\"\n lst = list(range(-99, 100, 2))\n odd_sum = sum([i**2 for i in lst if i%2!=0 and i > 0])\n assert candidate(lst) == odd_sum , \"This prints if this assert fails 7 (good for debugging!)\"\n\n", "entry_point": "double_the_difference"} {"task": "code_en_3511b0921b939c18c5b6fa70", "id": "3511b0921b939c18c5b6fa70", "domain": "code", "language": "en", "split": "validation", "group": "HumanEval/105", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef by_length(arr):\n \"\"\"\n Given an array of integers, sort the integers that are between 1 and 9 inclusive,\n reverse the resulting array, and then replace each digit by its corresponding name from\n \"One\", \"Two\", \"Three\", \"Four\", \"Five\", \"Six\", \"Seven\", \"Eight\", \"Nine\".\n\n For example:\n arr = [2, 1, 1, 4, 5, 8, 2, 3] \n -> sort arr -> [1, 1, 2, 2, 3, 4, 5, 8] \n -> reverse arr -> [8, 5, 4, 3, 2, 2, 1, 1]\n return [\"Eight\", \"Five\", \"Four\", \"Three\", \"Two\", \"Two\", \"One\", \"One\"]\n \n If the array is empty, return an empty array:\n arr = []\n return []\n \n If the array has any strange number ignore it:\n arr = [1, -1 , 55] \n -> sort arr -> [-1, 1, 55]\n -> reverse arr -> [55, 1, -1]\n return = ['One']\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly filters to integers 1–9 inclusive, sorts them in descending order (sort then reverse, or equivalent), and maps each to its name \"One\"…\"Nine\", producing [\"Eight\",\"Five\",\"Four\",\"Three\",\"Two\",\"Two\",\"One\",\"One\"] for [2,1,1,4,5,8,2,3]."}, {"id": "c2", "description": "Handles boundary cases: an empty list returns []; values outside the range such as 0, negatives, 10 and large numbers like 55 are ignored without crashing (e.g. [1,-1,55] -> ['One']); duplicates are preserved; a list with no valid digits returns []."}, {"id": "c3", "description": "The returned body is syntactically valid, consistently indented as a function body, has no repeated signature or Markdown, and terminates. A crash on any in-domain integer list, such as a KeyError or IndexError from mapping out-of-range values, scores 6 or below."}, {"id": "c4", "description": "Uses an appropriate, readable approach, such as a lookup list or dict plus filter and sort, with reasonable efficiency and no extra guards that change outputs."}]}, "target_words": null, "reference": " dic = {\n 1: \"One\",\n 2: \"Two\",\n 3: \"Three\",\n 4: \"Four\",\n 5: \"Five\",\n 6: \"Six\",\n 7: \"Seven\",\n 8: \"Eight\",\n 9: \"Nine\",\n }\n sorted_arr = sorted(arr, reverse=True)\n new_arr = []\n for var in sorted_arr:\n try:\n new_arr.append(dic[var])\n except:\n pass\n return new_arr\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/105"}, "prefix": "\ndef by_length(arr):\n \"\"\"\n Given an array of integers, sort the integers that are between 1 and 9 inclusive,\n reverse the resulting array, and then replace each digit by its corresponding name from\n \"One\", \"Two\", \"Three\", \"Four\", \"Five\", \"Six\", \"Seven\", \"Eight\", \"Nine\".\n\n For example:\n arr = [2, 1, 1, 4, 5, 8, 2, 3] \n -> sort arr -> [1, 1, 2, 2, 3, 4, 5, 8] \n -> reverse arr -> [8, 5, 4, 3, 2, 2, 1, 1]\n return [\"Eight\", \"Five\", \"Four\", \"Three\", \"Two\", \"Two\", \"One\", \"One\"]\n \n If the array is empty, return an empty array:\n arr = []\n return []\n \n If the array has any strange number ignore it:\n arr = [1, -1 , 55] \n -> sort arr -> [-1, 1, 55]\n -> reverse arr -> [55, 1, -1]\n return = ['One']\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([2, 1, 1, 4, 5, 8, 2, 3]) == [\"Eight\", \"Five\", \"Four\", \"Three\", \"Two\", \"Two\", \"One\", \"One\"], \"Error\"\n assert candidate([]) == [], \"Error\"\n assert candidate([1, -1 , 55]) == ['One'], \"Error\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate([1, -1, 3, 2]) == [\"Three\", \"Two\", \"One\"]\n assert candidate([9, 4, 8]) == [\"Nine\", \"Eight\", \"Four\"]\n\n", "entry_point": "by_length"} {"task": "code_en_3605351078b55d4f61145891", "id": "3605351078b55d4f61145891", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/81", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef numerical_letter_grade(grades):\n \"\"\"It is the last week of the semester and the teacher has to give the grades\n to students. The teacher has been making her own algorithm for grading.\n The only problem is, she has lost the code she used for grading.\n She has given you a list of GPAs for some students and you have to write \n a function that can output a list of letter grades using the following table:\n GPA | Letter grade\n 4.0 A+\n > 3.7 A \n > 3.3 A- \n > 3.0 B+\n > 2.7 B \n > 2.3 B-\n > 2.0 C+\n > 1.7 C\n > 1.3 C-\n > 1.0 D+ \n > 0.7 D \n > 0.0 D-\n 0.0 E\n \n\n Example:\n grade_equation([4.0, 3, 1.7, 2, 3.5]) ==> ['A+', 'B', 'C-', 'C', 'A-']\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Threshold mapping is correct with strict '>' comparisons: exactly 4.0 maps to 'A+', and boundary values map to the lower band (e.g., 3.7→'A-', 3.0→'B', 1.7→'C-', 2.0→'C', 1.0→'D', 0.7→'D-'). 0.0 maps to 'E', and any value in (0, 0.7] maps to 'D-'. The docstring example [4.0, 3, 1.7, 2, 3.5] yields ['A+', 'B', 'C-', 'C', 'A-']."}, {"id": "c2", "description": "Edge inputs are handled without crashing: an empty list returns [], integer inputs such as 3 or 0 work like floats, and output order and length match the input list. Values above 4.0 being treated as 'A+' is acceptable but not required; a crash on in-domain values (0.0–4.0) is severe."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation that fits under the given signature. It returns a list of strings rather than printing, and it does not repeat the signature or use Markdown."}, {"id": "c4", "description": "The logic is clear and linear, using an if/elif chain or an ordered threshold table, with no output-altering extra guards or needless complexity."}]}, "target_words": null, "reference": "\n \n letter_grade = []\n for gpa in grades:\n if gpa == 4.0:\n letter_grade.append(\"A+\")\n elif gpa > 3.7:\n letter_grade.append(\"A\")\n elif gpa > 3.3:\n letter_grade.append(\"A-\")\n elif gpa > 3.0:\n letter_grade.append(\"B+\")\n elif gpa > 2.7:\n letter_grade.append(\"B\")\n elif gpa > 2.3:\n letter_grade.append(\"B-\")\n elif gpa > 2.0:\n letter_grade.append(\"C+\")\n elif gpa > 1.7:\n letter_grade.append(\"C\")\n elif gpa > 1.3:\n letter_grade.append(\"C-\")\n elif gpa > 1.0:\n letter_grade.append(\"D+\")\n elif gpa > 0.7:\n letter_grade.append(\"D\")\n elif gpa > 0.0:\n letter_grade.append(\"D-\")\n else:\n letter_grade.append(\"E\")\n return letter_grade\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/81"}, "prefix": "\ndef numerical_letter_grade(grades):\n \"\"\"It is the last week of the semester and the teacher has to give the grades\n to students. The teacher has been making her own algorithm for grading.\n The only problem is, she has lost the code she used for grading.\n She has given you a list of GPAs for some students and you have to write \n a function that can output a list of letter grades using the following table:\n GPA | Letter grade\n 4.0 A+\n > 3.7 A \n > 3.3 A- \n > 3.0 B+\n > 2.7 B \n > 2.3 B-\n > 2.0 C+\n > 1.7 C\n > 1.3 C-\n > 1.0 D+ \n > 0.7 D \n > 0.0 D-\n 0.0 E\n \n\n Example:\n grade_equation([4.0, 3, 1.7, 2, 3.5]) ==> ['A+', 'B', 'C-', 'C', 'A-']\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([4.0, 3, 1.7, 2, 3.5]) == ['A+', 'B', 'C-', 'C', 'A-']\n assert candidate([1.2]) == ['D+']\n assert candidate([0.5]) == ['D-']\n assert candidate([0.0]) == ['E']\n assert candidate([1, 0.3, 1.5, 2.8, 3.3]) == ['D', 'D-', 'C-', 'B', 'B+']\n assert candidate([0, 0.7]) == ['E', 'D-']\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "numerical_letter_grade"} {"task": "code_en_3f22fb9def1ceba73dbeda84", "id": "3f22fb9def1ceba73dbeda84", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/0", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\nfrom typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n \"\"\" Check if in given list of numbers, are any two numbers closer to each other than\n given threshold.\n >>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n False\n >>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n True\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Functional correctness: the body returns True exactly when some pair of distinct-index elements has absolute difference strictly less than threshold, and False otherwise. Both docstring examples are reproduced (False for [1.0,2.0,3.0],0.5; True for the 2.8/3.0 case). A non-strict comparison (<=), comparing an element with itself, or using signed rather than absolute difference is a defect."}, {"id": "c2", "description": "Boundary cases: an empty list and a single-element list return False without crashing. Duplicate values (difference 0) return True for positive threshold. A pair whose difference equals the threshold exactly returns False. Negative numbers are handled via absolute difference. Threshold 0 or a negative threshold returns False, since no pair can have absolute difference below it. A crash or non-termination on any of these inputs means a score of 6 or below."}, {"id": "c3", "description": "Algorithmic suitability: the approach is either O(n log n) sort-and-compare-adjacent or a correct O(n^2) pairwise check. Sorting does not mutate the caller's list in a way that matters, or uses sorted(). There are no unnecessary guards that alter outputs. Any correct algorithm receives full credit."}, {"id": "c4", "description": "Validity and readability: the output is only the function body, without a repeated signature, docstring, or Markdown fences. It is consistently indented so it parses when placed under the given signature, including the first line. It uses clear names and has no extraneous commentary."}]}, "target_words": null, "reference": " for idx, elem in enumerate(numbers):\n for idx2, elem2 in enumerate(numbers):\n if idx != idx2:\n distance = abs(elem - elem2)\n if distance < threshold:\n return True\n\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/0"}, "prefix": "from typing import List\n\n\ndef has_close_elements(numbers: List[float], threshold: float) -> bool:\n \"\"\" Check if in given list of numbers, are any two numbers closer to each other than\n given threshold.\n >>> has_close_elements([1.0, 2.0, 3.0], 0.5)\n False\n >>> has_close_elements([1.0, 2.8, 3.0, 4.0, 5.0, 2.0], 0.3)\n True\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate([1.0, 2.0, 3.9, 4.0, 5.0, 2.2], 0.3) == True\n assert candidate([1.0, 2.0, 3.9, 4.0, 5.0, 2.2], 0.05) == False\n assert candidate([1.0, 2.0, 5.9, 4.0, 5.0], 0.95) == True\n assert candidate([1.0, 2.0, 5.9, 4.0, 5.0], 0.8) == False\n assert candidate([1.0, 2.0, 3.0, 4.0, 5.0, 2.0], 0.1) == True\n assert candidate([1.1, 2.2, 3.1, 4.1, 5.1], 1.0) == True\n assert candidate([1.1, 2.2, 3.1, 4.1, 5.1], 0.5) == False\n\n", "entry_point": "has_close_elements"} {"task": "code_en_40d9e23359615c32f3e133cb", "id": "40d9e23359615c32f3e133cb", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/126", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef is_sorted(lst):\n '''\n Given a list of numbers, return whether or not they are sorted\n in ascending order. If list has more than 1 duplicate of the same\n number, return False. Assume no negative numbers and only integers.\n\n Examples\n is_sorted([5]) ➞ True\n is_sorted([1, 2, 3, 4, 5]) ➞ True\n is_sorted([1, 3, 2, 4, 5]) ➞ False\n is_sorted([1, 2, 3, 4, 5, 6]) ➞ True\n is_sorted([1, 2, 3, 4, 5, 6, 7]) ➞ True\n is_sorted([1, 3, 2, 4, 5, 6, 7]) ➞ False\n is_sorted([1, 2, 2, 3, 3, 4]) ➞ True\n is_sorted([1, 2, 2, 2, 3, 4]) ➞ False\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly returns True only when the list is non-decreasing (adjacent equal values allowed), and returns False for any descending adjacent pair; matches all docstring examples, e.g. [1,3,2,4,5] -> False, [1,2,2,3,3,4] -> True."}, {"id": "c2", "description": "Correctly applies the duplicate rule: returns False when any number appears 3 or more times (e.g. [1,2,2,2,3,4] -> False), while exactly two occurrences are allowed. Counting is per value, not a global duplicate check."}, {"id": "c3", "description": "Handles boundary inputs without crashing or looping forever: the empty list returns True, a single element returns True, and lists containing 0 work correctly. A crash or non-termination on in-domain input scores 6 or below."}, {"id": "c4", "description": "The returned body is syntactically valid, properly indented Python with no signature or Markdown. It uses a reasonable algorithm, roughly O(n) or O(n log n), and is readable without extra guards that change outputs."}]}, "target_words": null, "reference": " count_digit = dict([(i, 0) for i in lst])\n for i in lst:\n count_digit[i]+=1 \n if any(count_digit[i] > 2 for i in lst):\n return False\n if all(lst[i-1] <= lst[i] for i in range(1, len(lst))):\n return True\n else:\n return False\n \n \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/126"}, "prefix": "\ndef is_sorted(lst):\n '''\n Given a list of numbers, return whether or not they are sorted\n in ascending order. If list has more than 1 duplicate of the same\n number, return False. Assume no negative numbers and only integers.\n\n Examples\n is_sorted([5]) ➞ True\n is_sorted([1, 2, 3, 4, 5]) ➞ True\n is_sorted([1, 3, 2, 4, 5]) ➞ False\n is_sorted([1, 2, 3, 4, 5, 6]) ➞ True\n is_sorted([1, 2, 3, 4, 5, 6, 7]) ➞ True\n is_sorted([1, 3, 2, 4, 5, 6, 7]) ➞ False\n is_sorted([1, 2, 2, 3, 3, 4]) ➞ True\n is_sorted([1, 2, 2, 2, 3, 4]) ➞ False\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([5]) == True\n assert candidate([1, 2, 3, 4, 5]) == True\n assert candidate([1, 3, 2, 4, 5]) == False\n assert candidate([1, 2, 3, 4, 5, 6]) == True\n assert candidate([1, 2, 3, 4, 5, 6, 7]) == True\n assert candidate([1, 3, 2, 4, 5, 6, 7]) == False, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([]) == True, \"This prints if this assert fails 2 (good for debugging!)\"\n assert candidate([1]) == True, \"This prints if this assert fails 3 (good for debugging!)\"\n assert candidate([3, 2, 1]) == False, \"This prints if this assert fails 4 (good for debugging!)\"\n \n # Check some edge cases that are easy to work out by hand.\n assert candidate([1, 2, 2, 2, 3, 4]) == False, \"This prints if this assert fails 5 (good for debugging!)\"\n assert candidate([1, 2, 3, 3, 3, 4]) == False, \"This prints if this assert fails 6 (good for debugging!)\"\n assert candidate([1, 2, 2, 3, 3, 4]) == True, \"This prints if this assert fails 7 (good for debugging!)\"\n assert candidate([1, 2, 3, 4]) == True, \"This prints if this assert fails 8 (good for debugging!)\"\n\n", "entry_point": "is_sorted"} {"task": "code_en_487545262fc4af3868fc55e8", "id": "487545262fc4af3868fc55e8", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/55", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef fib(n: int):\n \"\"\"Return n-th Fibonacci number.\n >>> fib(10)\n 55\n >>> fib(1)\n 1\n >>> fib(8)\n 21\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields fib(10)==55, fib(1)==1, fib(8)==21 under the standard indexing F(0)=0, F(1)=1, F(2)=1. Any mismatch on these examples is a major failure."}, {"id": "c2", "description": "Boundary handling: fib(0) returns 0 and fib(2) returns 1, with no crash, off-by-one error or infinite loop. Negative n is outside the stated domain and should not be heavily penalized unless it causes non-termination. Extra guards that change outputs for valid n should be penalized."}, {"id": "c3", "description": "Algorithmic suitability: the body uses an iterative, memoized or closed-form-safe integer approach that runs efficiently for moderately large n (e.g., n=100 or n=1000) and returns the exact integer. Penalize naive exponential recursion that effectively hangs, recursion-depth crashes on large n, and float-based Binet formulas that lose precision beyond roughly n=70. Treat a hang or crash as severe and score 6 or below."}, {"id": "c4", "description": "Validity and readability: the output is only the function body, with no signature, docstring or Markdown fences. It has consistent indentation, starting with the first line, so that it parses as a valid function body. Variable names and flow are clear. Lower this score if the indentation breaks the body."}]}, "target_words": null, "reference": " if n == 0:\n return 0\n if n == 1:\n return 1\n return fib(n - 1) + fib(n - 2)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/55"}, "prefix": "\n\ndef fib(n: int):\n \"\"\"Return n-th Fibonacci number.\n >>> fib(10)\n 55\n >>> fib(1)\n 1\n >>> fib(8)\n 21\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(10) == 55\n assert candidate(1) == 1\n assert candidate(8) == 21\n assert candidate(11) == 89\n assert candidate(12) == 144\n\n", "entry_point": "fib"} {"task": "code_en_4b664fa8ccb70c26e2b5ad47", "id": "4b664fa8ccb70c26e2b5ad47", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/147", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef get_max_triples(n):\n \"\"\"\n You are given a positive integer n. You have to create an integer array a of length n.\n For each i (1 ≤ i ≤ n), the value of a[i] = i * i - i + 1.\n Return the number of triples (a[i], a[j], a[k]) of a where i < j < k, \n and a[i] + a[j] + a[k] is a multiple of 3.\n\n Example :\n Input: n = 5\n Output: 1\n Explanation: \n a = [1, 3, 7, 13, 21]\n The only valid triple is (1, 7, 13).\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns the count of index triples i= 8 (the smallest sum of four positive even numbers is 2+2+2+2=8, and every larger even number is reachable). Mentally executing it must give False for 4 and 6 and True for 8, matching the docstring."}, {"id": "c2", "description": "Handles boundary and adversarial inputs correctly. It returns False for 0, for negative values such as -8, for odd values such as 7, 9 and 11, and for 2. It returns True for 10 and for large even values such as 10**18, without crashing or hanging."}, {"id": "c3", "description": "Uses a constant-time check such as `n % 2 == 0 and n >= 8`, or any other correct method that terminates quickly even on large n. A loop or enumeration that would be infeasibly slow, or would never finish, on large inputs is penalized as a severe failure."}, {"id": "c4", "description": "The returned body is syntactically valid Python with consistent indentation that fits under the given signature. It contains no Markdown fences and no repeated signature, returns a boolean, and is clear and concise. Extra guards that change outputs are not rewarded."}]}, "target_words": null, "reference": " return n%2 == 0 and n >= 8\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/138"}, "prefix": "\ndef is_equal_to_sum_even(n):\n \"\"\"Evaluate whether the given number n can be written as the sum of exactly 4 positive even numbers\n Example\n is_equal_to_sum_even(4) == False\n is_equal_to_sum_even(6) == False\n is_equal_to_sum_even(8) == True\n \"\"\"\n", "tests": "def check(candidate):\n assert candidate(4) == False\n assert candidate(6) == False\n assert candidate(8) == True\n assert candidate(10) == True\n assert candidate(11) == False\n assert candidate(12) == True\n assert candidate(13) == False\n assert candidate(16) == True\n", "entry_point": "is_equal_to_sum_even"} {"task": "code_en_518559926d019efe4a721054", "id": "518559926d019efe4a721054", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/128", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef prod_signs(arr):\n \"\"\"\n You are given an array arr of integers and you need to return\n sum of magnitudes of integers multiplied by product of all signs\n of each number in the array, represented by 1, -1 or 0.\n Note: return None for empty arr.\n\n Example:\n >>> prod_signs([1, 2, 2, -4]) == -9\n >>> prod_signs([0, 1]) == 0\n >>> prod_signs([]) == None\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns None for an empty list. For non-empty input it returns sum(abs(x)) multiplied by the product of signs (1, -1 or 0). This gives -9 for [1,2,2,-4] and 0 for [0,1]. Any zero in the list yields 0, and an even count of negatives yields a positive result, e.g. [-1,-2] gives 3. An empty list must not crash or return 0. If it does, score at most 6."}, {"id": "c2", "description": "Handles boundary inputs correctly: a single element (e.g. [-5] gives -5, [0] gives 0), all negatives with an odd count, multiple zeros, and large magnitudes, using exact integer arithmetic with no float conversion."}, {"id": "c3", "description": "The returned body is syntactically valid Python, consistently indented as a function body, and contains no signature, docstring repetition, or Markdown fences. Indentation errors that break the body lower this criterion."}, {"id": "c4", "description": "The algorithm is linear and readable, e.g. it counts negatives or detects zero instead of doing unnecessary work. Any correct approach gets full credit. Do not reward extra commentary or guards that change outputs."}]}, "target_words": null, "reference": " if not arr: return None\n prod = 0 if 0 in arr else (-1) ** len(list(filter(lambda x: x < 0, arr)))\n return prod * sum([abs(i) for i in arr])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/128"}, "prefix": "\ndef prod_signs(arr):\n \"\"\"\n You are given an array arr of integers and you need to return\n sum of magnitudes of integers multiplied by product of all signs\n of each number in the array, represented by 1, -1 or 0.\n Note: return None for empty arr.\n\n Example:\n >>> prod_signs([1, 2, 2, -4]) == -9\n >>> prod_signs([0, 1]) == 0\n >>> prod_signs([]) == None\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([1, 2, 2, -4]) == -9\n assert candidate([0, 1]) == 0\n assert candidate([1, 1, 1, 2, 3, -1, 1]) == -10\n assert candidate([]) == None\n assert candidate([2, 4,1, 2, -1, -1, 9]) == 20\n assert candidate([-1, 1, -1, 1]) == 4\n assert candidate([-1, 1, 1, 1]) == -4\n assert candidate([-1, 1, 1, 0]) == 0\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "prod_signs"} {"task": "code_en_566a680c31d0cc72151e8bf4", "id": "566a680c31d0cc72151e8bf4", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/39", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef prime_fib(n: int):\n \"\"\"\n prime_fib returns n-th number that is a Fibonacci number and it's also prime.\n >>> prime_fib(1)\n 2\n >>> prime_fib(2)\n 3\n >>> prime_fib(3)\n 5\n >>> prime_fib(4)\n 13\n >>> prime_fib(5)\n 89\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on the docstring examples: mentally executing the body gives prime_fib(1)=2, prime_fib(2)=3, prime_fib(3)=5, prime_fib(4)=13, prime_fib(5)=89. It also gives later values such as prime_fib(6)=233, prime_fib(7)=1597, prime_fib(8)=28657 and prime_fib(9)=514229. The Fibonacci sequence counts the value 1 only once or not at all, so no duplicate is produced. A wrong value for any docstring example scores this criterion low."}, {"id": "c2", "description": "Primality testing is correct. 0 and 1 are rejected as non-prime, 2 is accepted, and composite Fibonacci numbers such as 8, 21, 34, 55, 144 and 377 (=13*29) are rejected. The test checks divisors up to sqrt(k) inclusive, so perfect squares and products of two equal primes are handled correctly. An off-by-one in the divisor bound that misclassifies a composite scores this moderately low."}, {"id": "c3", "description": "Termination and robustness: for valid n>=1 the loop terminates and returns the n-th prime Fibonacci number without crashing, for example with no IndexError and no infinite loop. Primality uses trial division bounded by sqrt, or another correct method, so it stays practical for n up to about 10. Unbounded behavior on n<=0 is not penalized, because the docstring domain is n>=1. Non-termination or a crash for any n>=1 scores 6 or below."}, {"id": "c4", "description": "Validity and readability: the returned text is only an indented function body with consistent indentation, starting at the body level so it compiles under the given signature. It contains no repeated signature and no Markdown fences. Helper logic is clear. Broken first-line indentation or a syntax error lowers this criterion significantly."}]}, "target_words": null, "reference": " import math\n\n def is_prime(p):\n if p < 2:\n return False\n for k in range(2, min(int(math.sqrt(p)) + 1, p - 1)):\n if p % k == 0:\n return False\n return True\n f = [0, 1]\n while True:\n f.append(f[-1] + f[-2])\n if is_prime(f[-1]):\n n -= 1\n if n == 0:\n return f[-1]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/39"}, "prefix": "\n\ndef prime_fib(n: int):\n \"\"\"\n prime_fib returns n-th number that is a Fibonacci number and it's also prime.\n >>> prime_fib(1)\n 2\n >>> prime_fib(2)\n 3\n >>> prime_fib(3)\n 5\n >>> prime_fib(4)\n 13\n >>> prime_fib(5)\n 89\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(1) == 2\n assert candidate(2) == 3\n assert candidate(3) == 5\n assert candidate(4) == 13\n assert candidate(5) == 89\n assert candidate(6) == 233\n assert candidate(7) == 1597\n assert candidate(8) == 28657\n assert candidate(9) == 514229\n assert candidate(10) == 433494437\n\n", "entry_point": "prime_fib"} {"task": "code_en_5cd7e90044090a823eec1882", "id": "5cd7e90044090a823eec1882", "domain": "code", "language": "en", "split": "validation", "group": "HumanEval/144", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef simplify(x, n):\n \"\"\"Your task is to implement a function that will simplify the expression\n x * n. The function returns True if x * n evaluates to a whole number and False\n otherwise. Both x and n, are string representation of a fraction, and have the following format,\n / where both numerator and denominator are positive whole numbers.\n\n You can assume that x, and n are valid fractions, and do not have zero as denominator.\n\n simplify(\"1/5\", \"5/1\") = True\n simplify(\"1/6\", \"2/1\") = False\n simplify(\"7/10\", \"10/2\") = False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: parses both fraction strings into integer numerators/denominators and returns True exactly when (a_num*b_num) is divisible by (a_den*b_den), returning a boolean. Mentally verify the docstring examples (\"1/5\",\"5/1\")->True, (\"1/6\",\"2/1\")->False, (\"7/10\",\"10/2\")->False, plus cases like (\"2/4\",\"2/1\")->True and (\"3/7\",\"7/3\")->True. Wrong results on common inputs score low."}, {"id": "c2", "description": "Boundary and robustness: uses exact integer arithmetic (modulo or Fraction) rather than float division, so large numerators/denominators do not cause precision errors. Handles numerator 1 and denominator 1 correctly. Adds no guards that change outputs for valid positive fractions. Float-based checks that can misjudge large values are moderate failures. Crashes on valid inputs score 6 or below."}, {"id": "c3", "description": "Validity: the returned body is properly indented, with a consistent first line, and is syntactically valid Python that executes as the function body. It contains no repeated signature or Markdown fences. Indentation errors that break the body lower this criterion substantially."}, {"id": "c4", "description": "Readability and suitability: the solution is concise and clear, with an O(1)/O(digits) approach and meaningful variable names. Do not reward extra commentary or added complexity. Any correct alternative algorithm, such as using fractions.Fraction or gcd, earns full credit."}]}, "target_words": null, "reference": " a, b = x.split(\"/\")\n c, d = n.split(\"/\")\n numerator = int(a) * int(c)\n denom = int(b) * int(d)\n if (numerator/denom == int(numerator/denom)):\n return True\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/144"}, "prefix": "\ndef simplify(x, n):\n \"\"\"Your task is to implement a function that will simplify the expression\n x * n. The function returns True if x * n evaluates to a whole number and False\n otherwise. Both x and n, are string representation of a fraction, and have the following format,\n / where both numerator and denominator are positive whole numbers.\n\n You can assume that x, and n are valid fractions, and do not have zero as denominator.\n\n simplify(\"1/5\", \"5/1\") = True\n simplify(\"1/6\", \"2/1\") = False\n simplify(\"7/10\", \"10/2\") = False\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"1/5\", \"5/1\") == True, 'test1'\n assert candidate(\"1/6\", \"2/1\") == False, 'test2'\n assert candidate(\"5/1\", \"3/1\") == True, 'test3'\n assert candidate(\"7/10\", \"10/2\") == False, 'test4'\n assert candidate(\"2/10\", \"50/10\") == True, 'test5'\n assert candidate(\"7/2\", \"4/2\") == True, 'test6'\n assert candidate(\"11/6\", \"6/1\") == True, 'test7'\n assert candidate(\"2/3\", \"5/2\") == False, 'test8'\n assert candidate(\"5/2\", \"3/5\") == False, 'test9'\n assert candidate(\"2/4\", \"8/4\") == True, 'test10'\n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(\"2/4\", \"4/2\") == True, 'test11'\n assert candidate(\"1/5\", \"5/1\") == True, 'test12'\n assert candidate(\"1/5\", \"1/5\") == False, 'test13'\n\n", "entry_point": "simplify"} {"task": "code_en_5f545a140dfc262c9bec809d", "id": "5f545a140dfc262c9bec809d", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/59", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef largest_prime_factor(n: int):\n \"\"\"Return the largest prime factor of n. Assume n > 1 and is not a prime.\n >>> largest_prime_factor(13195)\n 29\n >>> largest_prime_factor(2048)\n 2\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: mentally executing the body returns 29 for 13195 and 2 for 2048, and returns the largest prime factor for other composite n > 1. Test perfect powers (e.g. 2048, 27 -> 3), squares of primes (49 -> 7), products with a large prime factor (e.g. 2*97 -> 97) and n whose largest factor remains after trial division (e.g. 15 -> 5). A crash or non-termination on any composite n > 1 is severe (score 6 or below). Inputs outside the domain (n <= 1 or prime n) need not be handled."}, {"id": "c2", "description": "Algorithmic suitability: the solution terminates in reasonable time for large composite n with a large prime factor. Trial division up to sqrt(n) while dividing out factors is preferred over iterating up to n or testing primality of every divisor naively. Any correct algorithm gets full credit, and this criterion does not override correctness."}, {"id": "c3", "description": "Validity of the returned body: it is only the function body, with no signature, docstring, or Markdown fences. Indentation is consistent, including the first line, so it parses when placed under the given signature. Every path returns an int, and the body does not add guards that change outputs for valid inputs."}, {"id": "c4", "description": "Readability: variable names are clear and the loop and termination logic are easy to follow, with no extraneous commentary or dead code. This criterion is weighted below correctness."}]}, "target_words": null, "reference": " def is_prime(k):\n if k < 2:\n return False\n for i in range(2, k - 1):\n if k % i == 0:\n return False\n return True\n largest = 1\n for j in range(2, n + 1):\n if n % j == 0 and is_prime(j):\n largest = max(largest, j)\n return largest\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/59"}, "prefix": "\n\ndef largest_prime_factor(n: int):\n \"\"\"Return the largest prime factor of n. Assume n > 1 and is not a prime.\n >>> largest_prime_factor(13195)\n 29\n >>> largest_prime_factor(2048)\n 2\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(15) == 5\n assert candidate(27) == 3\n assert candidate(63) == 7\n assert candidate(330) == 11\n assert candidate(13195) == 29\n\n", "entry_point": "largest_prime_factor"} {"task": "code_en_65d5c2bd2f60d3b82d981c90", "id": "65d5c2bd2f60d3b82d981c90", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/80", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef is_happy(s):\n \"\"\"You are given a string s.\n Your task is to check if the string is happy or not.\n A string is happy if its length is at least 3 and every 3 consecutive letters are distinct\n For example:\n is_happy(a) => False\n is_happy(aa) => False\n is_happy(abcd) => True\n is_happy(aabb) => False\n is_happy(adb) => True\n is_happy(xyy) => False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns False for any string with length less than 3 (e.g., '', 'a', 'aa') without crashing or raising an exception."}, {"id": "c2", "description": "For strings of length >= 3, checks every window of 3 consecutive characters (indices i, i+1, i+2 for all valid i, including the final window) and returns False if any two characters in a window are equal (s[i]==s[i+1], s[i+1]==s[i+2], or s[i]==s[i+2]). Otherwise returns True. Matches the examples: 'abcd'->True, 'aabb'->False, 'adb'->True, 'xyy'->False, and non-adjacent repeats like 'aba'->False."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation suitable for placement directly under the signature. It contains no Markdown fences and does not repeat the signature or docstring."}, {"id": "c4", "description": "Uses a linear-time approach (a single pass over the windows, or an equivalent) with clear logic. It has no off-by-one index errors that could cause an IndexError and no unnecessary complexity."}]}, "target_words": null, "reference": " if len(s) < 3:\n return False\n\n for i in range(len(s) - 2):\n \n if s[i] == s[i+1] or s[i+1] == s[i+2] or s[i] == s[i+2]:\n return False\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/80"}, "prefix": "\ndef is_happy(s):\n \"\"\"You are given a string s.\n Your task is to check if the string is happy or not.\n A string is happy if its length is at least 3 and every 3 consecutive letters are distinct\n For example:\n is_happy(a) => False\n is_happy(aa) => False\n is_happy(abcd) => True\n is_happy(aabb) => False\n is_happy(adb) => True\n is_happy(xyy) => False\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"a\") == False , \"a\"\n assert candidate(\"aa\") == False , \"aa\"\n assert candidate(\"abcd\") == True , \"abcd\"\n assert candidate(\"aabb\") == False , \"aabb\"\n assert candidate(\"adb\") == True , \"adb\"\n assert candidate(\"xyy\") == False , \"xyy\"\n assert candidate(\"iopaxpoi\") == True , \"iopaxpoi\"\n assert candidate(\"iopaxioi\") == False , \"iopaxioi\"\n", "entry_point": "is_happy"} {"task": "code_en_660ded474058ccc42d24efed", "id": "660ded474058ccc42d24efed", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/115", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef max_fill(grid, capacity):\n import math\n \"\"\"\n You are given a rectangular grid of wells. Each row represents a single well,\n and each 1 in a row represents a single unit of water.\n Each well has a corresponding bucket that can be used to extract water from it, \n and all buckets have the same capacity.\n Your task is to use the buckets to empty the wells.\n Output the number of times you need to lower the buckets.\n\n Example 1:\n Input: \n grid : [[0,0,1,0], [0,1,0,0], [1,1,1,1]]\n bucket_capacity : 1\n Output: 6\n\n Example 2:\n Input: \n grid : [[0,0,1,1], [0,0,0,0], [1,1,1,1], [0,1,1,1]]\n bucket_capacity : 2\n Output: 5\n \n Example 3:\n Input: \n grid : [[0,0,0], [0,0,0]]\n bucket_capacity : 5\n Output: 0\n\n Constraints:\n * all wells have the same length\n * 1 <= grid.length <= 10^2\n * 1 <= grid[:,1].length <= 10^2\n * grid[i][j] -> 0 | 1\n * 1 <= capacity <= 10\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Computes, for each row (well) independently, the ceiling of (number of 1s in the row / capacity) and returns the sum over all rows; produces 6, 5 and 0 on the three docstring examples. Pooling all water across wells before dividing (e.g., ceil(total/capacity)) is incorrect and should score low."}, {"id": "c2", "description": "Handles boundary cases correctly: rows of all zeros contribute 0 (not 1), capacity of 1 yields the total count of 1s, a row whose count is an exact multiple of capacity does not get an extra lowering, and capacity larger than a row's count gives 1 for nonempty rows. Uses exact integer ceiling (e.g., math.ceil(sum/cap) or (n+cap-1)//cap) without floating-point or off-by-one errors."}, {"id": "c3", "description": "Returned output is only a validly indented Python function body (no signature, docstring repetition, or Markdown fences) that runs without syntax or indentation errors and does not crash or loop on any input within the stated constraints."}, {"id": "c4", "description": "The approach is a simple linear pass over the grid (O(rows*cols)) and the code is concise and readable; no extraneous guards or behaviors that change outputs for valid inputs."}]}, "target_words": null, "reference": " return sum([math.ceil(sum(arr)/capacity) for arr in grid])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/115"}, "prefix": "\ndef max_fill(grid, capacity):\n import math\n \"\"\"\n You are given a rectangular grid of wells. Each row represents a single well,\n and each 1 in a row represents a single unit of water.\n Each well has a corresponding bucket that can be used to extract water from it, \n and all buckets have the same capacity.\n Your task is to use the buckets to empty the wells.\n Output the number of times you need to lower the buckets.\n\n Example 1:\n Input: \n grid : [[0,0,1,0], [0,1,0,0], [1,1,1,1]]\n bucket_capacity : 1\n Output: 6\n\n Example 2:\n Input: \n grid : [[0,0,1,1], [0,0,0,0], [1,1,1,1], [0,1,1,1]]\n bucket_capacity : 2\n Output: 5\n \n Example 3:\n Input: \n grid : [[0,0,0], [0,0,0]]\n bucket_capacity : 5\n Output: 0\n\n Constraints:\n * all wells have the same length\n * 1 <= grid.length <= 10^2\n * 1 <= grid[:,1].length <= 10^2\n * grid[i][j] -> 0 | 1\n * 1 <= capacity <= 10\n \"\"\"\n", "tests": "def check(candidate):\n\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([[0,0,1,0], [0,1,0,0], [1,1,1,1]], 1) == 6, \"Error\"\n assert candidate([[0,0,1,1], [0,0,0,0], [1,1,1,1], [0,1,1,1]], 2) == 5, \"Error\"\n assert candidate([[0,0,0], [0,0,0]], 5) == 0, \"Error\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate([[1,1,1,1], [1,1,1,1]], 2) == 4, \"Error\"\n assert candidate([[1,1,1,1], [1,1,1,1]], 9) == 2, \"Error\"\n\n", "entry_point": "max_fill"} {"task": "code_en_6c207c945f0e1d0eedef36ea", "id": "6c207c945f0e1d0eedef36ea", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/142", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\n\ndef sum_squares(lst):\n \"\"\"\"\n This function will take a list of integers. For all entries in the list, the function shall square the integer entry if its index is a \n multiple of 3 and will cube the integer entry if its index is a multiple of 4 and not a multiple of 3. The function will not \n change the entries in the list whose indexes are not a multiple of 3 or 4. The function shall then return the sum of all entries. \n \n Examples:\n For lst = [1,2,3] the output should be 6\n For lst = [] the output should be 0\n For lst = [-1,-5,2,-1,-5] the output should be -126\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Index rules are correct: the element at index i is squared when i % 3 == 0 (including index 0), cubed when i % 4 == 0 and i % 3 != 0, and left unchanged otherwise. Index 12, for example, is squared, not cubed. The code returns 6 for [1,2,3] and -126 for [-1,-5,2,-1,-5]."}, {"id": "c2", "description": "Boundary cases are handled: an empty list returns 0 without crashing. Negative values are handled with the correct sign, so cubes of negatives stay negative and squares are positive. Single-element lists and long lists sum correctly. The code does not raise an error on any list of integers."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation that fits under the given signature. Its first line is indented correctly. It contains no Markdown, and it does not repeat the signature or docstring. A body that breaks on paste should score lower here."}, {"id": "c4", "description": "The solution uses a single linear pass, such as enumerate with a sum or a loop. It reads clearly. It does not rely on unnecessary complexity or on guards that change outputs. Whether the input list is mutated in place does not affect the score as long as the returned sum is correct."}]}, "target_words": null, "reference": " result =[]\n for i in range(len(lst)):\n if i %3 == 0:\n result.append(lst[i]**2)\n elif i % 4 == 0 and i%3 != 0:\n result.append(lst[i]**3)\n else:\n result.append(lst[i])\n return sum(result)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/142"}, "prefix": "\n\n\ndef sum_squares(lst):\n \"\"\"\"\n This function will take a list of integers. For all entries in the list, the function shall square the integer entry if its index is a \n multiple of 3 and will cube the integer entry if its index is a multiple of 4 and not a multiple of 3. The function will not \n change the entries in the list whose indexes are not a multiple of 3 or 4. The function shall then return the sum of all entries. \n \n Examples:\n For lst = [1,2,3] the output should be 6\n For lst = [] the output should be 0\n For lst = [-1,-5,2,-1,-5] the output should be -126\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n \n assert candidate([1,2,3]) == 6\n assert candidate([1,4,9]) == 14\n assert candidate([]) == 0\n assert candidate([1,1,1,1,1,1,1,1,1]) == 9\n assert candidate([-1,-1,-1,-1,-1,-1,-1,-1,-1]) == -3\n assert candidate([0]) == 0\n assert candidate([-1,-5,2,-1,-5]) == -126\n assert candidate([-56,-99,1,0,-2]) == 3030\n assert candidate([-1,0,0,0,0,0,0,0,-1]) == 0\n assert candidate([-16, -9, -2, 36, 36, 26, -20, 25, -40, 20, -4, 12, -26, 35, 37]) == -14196\n assert candidate([-1, -3, 17, -1, -15, 13, -1, 14, -14, -12, -5, 14, -14, 6, 13, 11, 16, 16, 4, 10]) == -1448\n \n \n # Don't remove this line:\n", "entry_point": "sum_squares"} {"task": "code_en_6d8dcfb6bc0b3184f6539a8e", "id": "6d8dcfb6bc0b3184f6539a8e", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/29", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\nfrom typing import List\n\n\ndef filter_by_prefix(strings: List[str], prefix: str) -> List[str]:\n \"\"\" Filter an input list of strings only for ones that start with a given prefix.\n >>> filter_by_prefix([], 'a')\n []\n >>> filter_by_prefix(['abc', 'bcd', 'cde', 'array'], 'a')\n ['abc', 'array']\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns a new list containing exactly the strings that start with the prefix, e.g. via str.startswith or an equivalent slice comparison. It preserves the original order and duplicates. On the docstring examples it yields [] for an empty input and ['abc', 'array'] for the sample list."}, {"id": "c2", "description": "Boundary cases: an empty prefix returns every string. Strings shorter than the prefix are excluded without raising an error. An empty-string element is included only when the prefix is empty. Matching is case-sensitive. A crash or non-termination on any of these in-domain inputs scores 6 or below."}, {"id": "c3", "description": "Validity and readability: the output is only an indented function body with no signature and no Markdown. Indentation and syntax are valid as given. The approach is a simple linear pass, such as a list comprehension, with no guards that change the outputs."}]}, "target_words": null, "reference": " return [x for x in strings if x.startswith(prefix)]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/29"}, "prefix": "from typing import List\n\n\ndef filter_by_prefix(strings: List[str], prefix: str) -> List[str]:\n \"\"\" Filter an input list of strings only for ones that start with a given prefix.\n >>> filter_by_prefix([], 'a')\n []\n >>> filter_by_prefix(['abc', 'bcd', 'cde', 'array'], 'a')\n ['abc', 'array']\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate([], 'john') == []\n assert candidate(['xxx', 'asd', 'xxy', 'john doe', 'xxxAAA', 'xxx'], 'xxx') == ['xxx', 'xxxAAA', 'xxx']\n", "entry_point": "filter_by_prefix"} {"task": "code_en_76a9d2de7eb3c68b5aa9544d", "id": "76a9d2de7eb3c68b5aa9544d", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/69", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef search(lst):\n '''\n You are given a non-empty list of positive integers. Return the greatest integer that is greater than \n zero, and has a frequency greater than or equal to the value of the integer itself. \n The frequency of an integer is the number of times it appears in the list.\n If no such a value exist, return -1.\n Examples:\n search([4, 1, 2, 2, 3, 1]) == 2\n search([1, 2, 2, 3, 3, 3, 4, 4, 4]) == 3\n search([5, 5, 4, 4, 4]) == -1\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core examples: mentally executing the body yields 2 for [4,1,2,2,3,1], 3 for [1,2,2,3,3,3,4,4,4], and -1 for [5,5,4,4,4]. The body counts frequencies accurately and returns the greatest qualifying value, not the first or smallest one found. If any example fails, score this criterion low."}, {"id": "c2", "description": "Boundary cases within the stated domain of non-empty positive integers: [1] returns 1. [2] returns -1. [2,2] returns 2. [3,3,3,2,2] returns 3. A single large value such as [1000] returns -1 without crashing or allocating excessive memory, for example by sizing an array to max(lst)+1 with huge values. A crash or non-termination on a valid input scores 6 or below. Do not require handling of empty lists, zeros, or negatives."}, {"id": "c3", "description": "Syntactic validity: the returned text is only an indented function body, with no repeated signature, docstring, or Markdown fences. The indentation is consistent so that it parses when placed under the given def line. A first-line indentation mismatch or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: it uses a reasonable counting approach, such as Counter or a dict, at about O(n) or O(n log n). It avoids gratuitous O(n^2) repeated lst.count calls on large inputs, though those calls are only a minor deduction if the output is correct. The code is clear, and no extra guards change outputs for valid inputs."}]}, "target_words": null, "reference": " frq = [0] * (max(lst) + 1)\n for i in lst:\n frq[i] += 1;\n\n ans = -1\n for i in range(1, len(frq)):\n if frq[i] >= i:\n ans = i\n \n return ans\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/69"}, "prefix": "\ndef search(lst):\n '''\n You are given a non-empty list of positive integers. Return the greatest integer that is greater than \n zero, and has a frequency greater than or equal to the value of the integer itself. \n The frequency of an integer is the number of times it appears in the list.\n If no such a value exist, return -1.\n Examples:\n search([4, 1, 2, 2, 3, 1]) == 2\n search([1, 2, 2, 3, 3, 3, 4, 4, 4]) == 3\n search([5, 5, 4, 4, 4]) == -1\n '''\n", "tests": "def check(candidate):\n\n # manually generated tests\n assert candidate([5, 5, 5, 5, 1]) == 1\n assert candidate([4, 1, 4, 1, 4, 4]) == 4\n assert candidate([3, 3]) == -1\n assert candidate([8, 8, 8, 8, 8, 8, 8, 8]) == 8\n assert candidate([2, 3, 3, 2, 2]) == 2\n\n # automatically generated tests\n assert candidate([2, 7, 8, 8, 4, 8, 7, 3, 9, 6, 5, 10, 4, 3, 6, 7, 1, 7, 4, 10, 8, 1]) == 1\n assert candidate([3, 2, 8, 2]) == 2\n assert candidate([6, 7, 1, 8, 8, 10, 5, 8, 5, 3, 10]) == 1\n assert candidate([8, 8, 3, 6, 5, 6, 4]) == -1\n assert candidate([6, 9, 6, 7, 1, 4, 7, 1, 8, 8, 9, 8, 10, 10, 8, 4, 10, 4, 10, 1, 2, 9, 5, 7, 9]) == 1\n assert candidate([1, 9, 10, 1, 3]) == 1\n assert candidate([6, 9, 7, 5, 8, 7, 5, 3, 7, 5, 10, 10, 3, 6, 10, 2, 8, 6, 5, 4, 9, 5, 3, 10]) == 5\n assert candidate([1]) == 1\n assert candidate([8, 8, 10, 6, 4, 3, 5, 8, 2, 4, 2, 8, 4, 6, 10, 4, 2, 1, 10, 2, 1, 1, 5]) == 4\n assert candidate([2, 10, 4, 8, 2, 10, 5, 1, 2, 9, 5, 5, 6, 3, 8, 6, 4, 10]) == 2\n assert candidate([1, 6, 10, 1, 6, 9, 10, 8, 6, 8, 7, 3]) == 1\n assert candidate([9, 2, 4, 1, 5, 1, 5, 2, 5, 7, 7, 7, 3, 10, 1, 5, 4, 2, 8, 4, 1, 9, 10, 7, 10, 2, 8, 10, 9, 4]) == 4\n assert candidate([2, 6, 4, 2, 8, 7, 5, 6, 4, 10, 4, 6, 3, 7, 8, 8, 3, 1, 4, 2, 2, 10, 7]) == 4\n assert candidate([9, 8, 6, 10, 2, 6, 10, 2, 7, 8, 10, 3, 8, 2, 6, 2, 3, 1]) == 2\n assert candidate([5, 5, 3, 9, 5, 6, 3, 2, 8, 5, 6, 10, 10, 6, 8, 4, 10, 7, 7, 10, 8]) == -1\n assert candidate([10]) == -1\n assert candidate([9, 7, 7, 2, 4, 7, 2, 10, 9, 7, 5, 7, 2]) == 2\n assert candidate([5, 4, 10, 2, 1, 1, 10, 3, 6, 1, 8]) == 1\n assert candidate([7, 9, 9, 9, 3, 4, 1, 5, 9, 1, 2, 1, 1, 10, 7, 5, 6, 7, 6, 7, 7, 6]) == 1\n assert candidate([3, 10, 10, 9, 2]) == -1\n\n", "entry_point": "search"} {"task": "code_en_7c6fa249c4bc868decd3b20f", "id": "7c6fa249c4bc868decd3b20f", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/117", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef select_words(s, n):\n \"\"\"Given a string s and a natural number n, you have been tasked to implement \n a function that returns a list of all words from string s that contain exactly \n n consonants, in order these words appear in the string s.\n If the string s is empty then the function should return an empty list.\n Note: you may assume the input string contains only letters and spaces.\n Examples:\n select_words(\"Mary had a little lamb\", 4) ==> [\"little\"]\n select_words(\"Mary had a little lamb\", 3) ==> [\"Mary\", \"lamb\"]\n select_words(\"simple white space\", 2) ==> []\n select_words(\"Hello world\", 4) ==> [\"world\"]\n select_words(\"Uncle sam\", 3) ==> [\"Uncle\"]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Consonant counting is correct and case-insensitive: every letter not in a/e/i/o/u (upper or lower case) counts as a consonant, 'y' counts as a consonant, and vowels such as uppercase 'U' in \"Uncle\" are not counted. All five docstring examples produce the stated outputs, e.g. \"Mary had a little lamb\", 3 -> [\"Mary\", \"lamb\"]."}, {"id": "c2", "description": "Words with exactly n consonants are returned in their original order and original casing, and duplicates are preserved. The empty string returns []. Splitting handles multiple, leading or trailing spaces without producing empty-string words, and with n=0 an empty word is never returned. n=0 selects only all-vowel words. A crash or wrong result on these in-domain inputs lowers this criterion."}, {"id": "c3", "description": "The returned text is a valid, properly indented function body with no repeated signature, no docstring and no Markdown fences. The first line's indentation is consistent with the rest of the body, so it runs as given. It returns a list rather than a generator or a printed value."}, {"id": "c4", "description": "The approach is linear and readable, for example a split followed by a per-word count against a vowel set. It adds no extra guards or filtering that change outputs for valid inputs, and gains no credit for commentary."}]}, "target_words": null, "reference": " result = []\n for word in s.split():\n n_consonants = 0\n for i in range(0, len(word)):\n if word[i].lower() not in [\"a\",\"e\",\"i\",\"o\",\"u\"]:\n n_consonants += 1 \n if n_consonants == n:\n result.append(word)\n return result\n\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/117"}, "prefix": "\ndef select_words(s, n):\n \"\"\"Given a string s and a natural number n, you have been tasked to implement \n a function that returns a list of all words from string s that contain exactly \n n consonants, in order these words appear in the string s.\n If the string s is empty then the function should return an empty list.\n Note: you may assume the input string contains only letters and spaces.\n Examples:\n select_words(\"Mary had a little lamb\", 4) ==> [\"little\"]\n select_words(\"Mary had a little lamb\", 3) ==> [\"Mary\", \"lamb\"]\n select_words(\"simple white space\", 2) ==> []\n select_words(\"Hello world\", 4) ==> [\"world\"]\n select_words(\"Uncle sam\", 3) ==> [\"Uncle\"]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"Mary had a little lamb\", 4) == [\"little\"], \"First test error: \" + str(candidate(\"Mary had a little lamb\", 4)) \n assert candidate(\"Mary had a little lamb\", 3) == [\"Mary\", \"lamb\"], \"Second test error: \" + str(candidate(\"Mary had a little lamb\", 3)) \n assert candidate(\"simple white space\", 2) == [], \"Third test error: \" + str(candidate(\"simple white space\", 2)) \n assert candidate(\"Hello world\", 4) == [\"world\"], \"Fourth test error: \" + str(candidate(\"Hello world\", 4)) \n assert candidate(\"Uncle sam\", 3) == [\"Uncle\"], \"Fifth test error: \" + str(candidate(\"Uncle sam\", 3))\n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(\"\", 4) == [], \"1st edge test error: \" + str(candidate(\"\", 4))\n assert candidate(\"a b c d e f\", 1) == [\"b\", \"c\", \"d\", \"f\"], \"2nd edge test error: \" + str(candidate(\"a b c d e f\", 1))\n\n", "entry_point": "select_words"} {"task": "code_en_9274c55dd789316060304705", "id": "9274c55dd789316060304705", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/97", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef multiply(a, b):\n \"\"\"Complete the function that takes two integers and returns \n the product of their unit digits.\n Assume the input is always valid.\n Examples:\n multiply(148, 412) should return 16.\n multiply(19, 28) should return 72.\n multiply(2020, 1851) should return 0.\n multiply(14,-15) should return 20.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly computes the product of the unit digits for all docstring examples (148,412→16; 19,28→72; 2020,1851→0; 14,-15→20), extracting each unit digit as abs(n) % 10 (or equivalent) so negatives give a nonnegative digit; using raw n % 10 on negatives (e.g. -15 % 10 = 5 happens to work, but -14 % 10 = 6 is wrong) counts as a correctness failure."}, {"id": "c2", "description": "Handles boundary inputs correctly: zero for either argument (returns 0), single-digit numbers, both arguments negative (e.g. multiply(-3,-7) → 21), and very large integers, without crashing or producing wrong values."}, {"id": "c3", "description": "Returned body is syntactically valid, consistently indented as a function body (first line included), contains no signature/docstring repetition or Markdown, and returns an int."}, {"id": "c4", "description": "Uses a simple O(1) arithmetic approach (or an equally correct string-based one) that is readable, with no extra guards or special cases that alter the specified outputs."}]}, "target_words": null, "reference": " return abs(a % 10) * abs(b % 10)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/97"}, "prefix": "\ndef multiply(a, b):\n \"\"\"Complete the function that takes two integers and returns \n the product of their unit digits.\n Assume the input is always valid.\n Examples:\n multiply(148, 412) should return 16.\n multiply(19, 28) should return 72.\n multiply(2020, 1851) should return 0.\n multiply(14,-15) should return 20.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(148, 412) == 16, \"First test error: \" + str(candidate(148, 412)) \n assert candidate(19, 28) == 72, \"Second test error: \" + str(candidate(19, 28)) \n assert candidate(2020, 1851) == 0, \"Third test error: \" + str(candidate(2020, 1851))\n assert candidate(14,-15) == 20, \"Fourth test error: \" + str(candidate(14,-15)) \n assert candidate(76, 67) == 42, \"Fifth test error: \" + str(candidate(76, 67)) \n assert candidate(17, 27) == 49, \"Sixth test error: \" + str(candidate(17, 27)) \n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(0, 1) == 0, \"1st edge test error: \" + str(candidate(0, 1))\n assert candidate(0, 0) == 0, \"2nd edge test error: \" + str(candidate(0, 0))\n\n", "entry_point": "multiply"} {"task": "code_en_9900bb5e507943b17b8b9aa1", "id": "9900bb5e507943b17b8b9aa1", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/24", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef largest_divisor(n: int) -> int:\n \"\"\" For a given number n, find the largest number that divides n evenly, smaller than n\n >>> largest_divisor(15)\n 5\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core cases: the body returns the largest proper divisor of n (the largest d < n with n % d == 0). Mentally executing it should give largest_divisor(15)==5, (10)==5, (49)==7, (100)==50, (3)==1 for a prime, and (2)==1. Any wrong value on these typical inputs is a major failure."}, {"id": "c2", "description": "Boundary handling within the stated domain (positive integers n >= 2, with n=1 as a plausible edge): the code must not crash or loop forever. For example, there must be no modulo-by-zero from a range that reaches 0. Non-termination or an exception for n >= 2 scores 6 or below. For n=1 there is no proper divisor, so either returning 1/0 or raising is acceptable as long as the code terminates. No invented requirements for negative inputs are needed."}, {"id": "c3", "description": "Validity of the returned body: it is just an indented function body with no repeated signature, docstring or Markdown fences. Indentation is consistent, so it parses when placed under the given def, and it uses no undefined names. Problems with first-line indentation or syntax lower this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is reasonable, such as a descending scan from n-1 or n//2, or finding the smallest factor up to sqrt(n) and returning n//p. It returns early, and the logic is clear and concise. Any correct algorithm gets full credit, and extra guards that change outputs are penalized."}]}, "target_words": null, "reference": " for i in reversed(range(n)):\n if n % i == 0:\n return i\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/24"}, "prefix": "\n\ndef largest_divisor(n: int) -> int:\n \"\"\" For a given number n, find the largest number that divides n evenly, smaller than n\n >>> largest_divisor(15)\n 5\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate(3) == 1\n assert candidate(7) == 1\n assert candidate(10) == 5\n assert candidate(100) == 50\n assert candidate(49) == 7\n", "entry_point": "largest_divisor"} {"task": "code_en_9ee83c01914b9afaac1d30df", "id": "9ee83c01914b9afaac1d30df", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/112", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef reverse_delete(s,c):\n \"\"\"Task\n We are given two strings s and c, you have to deleted all the characters in s that are equal to any character in c\n then check if the result string is palindrome.\n A string is called palindrome if it reads the same backward as forward.\n You should return a tuple containing the result string and True/False for the check.\n Example\n For s = \"abcde\", c = \"ae\", the result should be ('bcd',False)\n For s = \"abcdef\", c = \"b\" the result should be ('acdef',False)\n For s = \"abcdedcba\", c = \"ab\", the result should be ('cdedc',True)\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly removes every character of s that appears anywhere in c (character-set membership, not substring removal), preserving the order of the remaining characters. Mentally running the code on the docstring examples yields 'bcd', 'acdef' and 'cdedc'."}, {"id": "c2", "description": "Returns a tuple (result_string, bool). The bool is True exactly when the filtered string equals its reverse, so the examples give False, False and True. A list, a swapped order, or a non-boolean truthy value is a fault."}, {"id": "c3", "description": "Handles boundary cases without crashing. An empty s gives ('', True). An empty c leaves s unchanged. When c removes every character, the result is ('', True). Repeated characters in c are handled, and the code makes no assumptions about case or character type beyond exact matching."}, {"id": "c4", "description": "The returned body is valid Python with consistent indentation that works as a function body, with no repeated signature and no Markdown. It uses a reasonable linear-time approach, such as set membership or join with a comprehension, and stays readable."}]}, "target_words": null, "reference": " s = ''.join([char for char in s if char not in c])\n return (s,s[::-1] == s)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/112"}, "prefix": "\ndef reverse_delete(s,c):\n \"\"\"Task\n We are given two strings s and c, you have to deleted all the characters in s that are equal to any character in c\n then check if the result string is palindrome.\n A string is called palindrome if it reads the same backward as forward.\n You should return a tuple containing the result string and True/False for the check.\n Example\n For s = \"abcde\", c = \"ae\", the result should be ('bcd',False)\n For s = \"abcdef\", c = \"b\" the result should be ('acdef',False)\n For s = \"abcdedcba\", c = \"ab\", the result should be ('cdedc',True)\n \"\"\"\n", "tests": "def check(candidate):\n\n assert candidate(\"abcde\",\"ae\") == ('bcd',False)\n assert candidate(\"abcdef\", \"b\") == ('acdef',False)\n assert candidate(\"abcdedcba\",\"ab\") == ('cdedc',True)\n assert candidate(\"dwik\",\"w\") == ('dik',False)\n assert candidate(\"a\",\"a\") == ('',True)\n assert candidate(\"abcdedcba\",\"\") == ('abcdedcba',True)\n assert candidate(\"abcdedcba\",\"v\") == ('abcdedcba',True)\n assert candidate(\"vabba\",\"v\") == ('abba',True)\n assert candidate(\"mamma\", \"mia\") == (\"\", True)\n", "entry_point": "reverse_delete"} {"task": "code_en_a67b2da0dbd5915b5ced9c07", "id": "a67b2da0dbd5915b5ced9c07", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/127", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef intersection(interval1, interval2):\n \"\"\"You are given two intervals,\n where each interval is a pair of integers. For example, interval = (start, end) = (1, 2).\n The given intervals are closed which means that the interval (start, end)\n includes both start and end.\n For each given interval, it is assumed that its start is less or equal its end.\n Your task is to determine whether the length of intersection of these two \n intervals is a prime number.\n Example, the intersection of the intervals (1, 3), (2, 4) is (2, 3)\n which its length is 1, which not a prime number.\n If the length of the intersection is a prime number, return \"YES\",\n otherwise, return \"NO\".\n If the two intervals don't intersect, return \"NO\".\n\n\n [input/output] samples:\n intersection((1, 2), (2, 3)) ==> \"NO\"\n intersection((-1, 1), (0, 4)) ==> \"NO\"\n intersection((-3, -1), (-5, 5)) ==> \"YES\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Computes intersection as (max of starts, min of ends) and length as end minus start (not +1), so (1,3),(2,4) gives length 1 -> \"NO\", (-3,-1),(-5,5) gives length 2 -> \"YES\", and (1,2),(2,3) gives length 0 -> \"NO\"; returns exact strings \"YES\"/\"NO\"."}, {"id": "c2", "description": "Handles boundary cases: disjoint intervals (negative length) and zero/one lengths return \"NO\" without crashing; the primality test correctly rejects 0, 1 and negatives and accepts 2 and 3; it works for negative coordinates and for identical or nested intervals."}, {"id": "c3", "description": "Primality check is correct and terminates efficiently for large lengths (e.g., trial division up to sqrt(n)); it does no infinite looping or excessive work on large inputs. Any correct alternative algorithm gets full credit."}, {"id": "c4", "description": "The returned body is valid, consistently indented Python, with no repeated signature or Markdown, and it runs as the function body. The logic is readable and has no extra guards that alter specified outputs."}]}, "target_words": null, "reference": " def is_prime(num):\n if num == 1 or num == 0:\n return False\n if num == 2:\n return True\n for i in range(2, num):\n if num%i == 0:\n return False\n return True\n\n l = max(interval1[0], interval2[0])\n r = min(interval1[1], interval2[1])\n length = r - l\n if length > 0 and is_prime(length):\n return \"YES\"\n return \"NO\"\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/127"}, "prefix": "\ndef intersection(interval1, interval2):\n \"\"\"You are given two intervals,\n where each interval is a pair of integers. For example, interval = (start, end) = (1, 2).\n The given intervals are closed which means that the interval (start, end)\n includes both start and end.\n For each given interval, it is assumed that its start is less or equal its end.\n Your task is to determine whether the length of intersection of these two \n intervals is a prime number.\n Example, the intersection of the intervals (1, 3), (2, 4) is (2, 3)\n which its length is 1, which not a prime number.\n If the length of the intersection is a prime number, return \"YES\",\n otherwise, return \"NO\".\n If the two intervals don't intersect, return \"NO\".\n\n\n [input/output] samples:\n intersection((1, 2), (2, 3)) ==> \"NO\"\n intersection((-1, 1), (0, 4)) ==> \"NO\"\n intersection((-3, -1), (-5, 5)) ==> \"YES\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate((1, 2), (2, 3)) == \"NO\"\n assert candidate((-1, 1), (0, 4)) == \"NO\"\n assert candidate((-3, -1), (-5, 5)) == \"YES\"\n assert candidate((-2, 2), (-4, 0)) == \"YES\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate((-11, 2), (-1, -1)) == \"NO\"\n assert candidate((1, 2), (3, 5)) == \"NO\"\n assert candidate((1, 2), (1, 2)) == \"NO\"\n assert candidate((-2, -2), (-3, -2)) == \"NO\"\n\n", "entry_point": "intersection"} {"task": "code_en_a9bcc72a33366fcfdbf5281a", "id": "a9bcc72a33366fcfdbf5281a", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/76", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef is_simple_power(x, n):\n \"\"\"Your task is to write a function that returns true if a number x is a simple\n power of n and false in other cases.\n x is a simple power of n if n**int=x\n For example:\n is_simple_power(1, 4) => true\n is_simple_power(2, 2) => true\n is_simple_power(8, 2) => true\n is_simple_power(3, 2) => false\n is_simple_power(3, 1) => false\n is_simple_power(5, 3) => false\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: returns True for (1,4), (2,2), (8,2) and False for (3,2), (3,1), (5,3). x=1 must return True for any n because n**0 == 1. Returns proper Python booleans or equivalent truthy/falsy values."}, {"id": "c2", "description": "Boundary handling and termination. n=1 must terminate and return True only when x==1, so (3,1) is False with no infinite loop. Inputs like n=0, n=-1, x=0, or negative x must not cause non-termination or a crash. A hang or exception on n=1 or x=1 is severe and scores 6 or below. Wrong values on rare inputs such as n=0 or negative bases are moderate deductions."}, {"id": "c3", "description": "Algorithmic suitability: uses exact integer arithmetic, such as repeated multiplication until the product reaches or exceeds x, or repeated division. It avoids floating-point log-based checks that misjudge large powers such as (3**20, 3) or (2**50, 2). It runs in logarithmic iterations and stays correct for large x."}, {"id": "c4", "description": "Validity and readability: the output is only a correctly indented function body that parses as valid Python, with no repeated signature and no Markdown. The logic is concise and clear, without extraneous commentary or guards that change outputs."}]}, "target_words": null, "reference": " if (n == 1): \n return (x == 1) \n power = 1\n while (power < x): \n power = power * n \n return (power == x) \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/76"}, "prefix": "\ndef is_simple_power(x, n):\n \"\"\"Your task is to write a function that returns true if a number x is a simple\n power of n and false in other cases.\n x is a simple power of n if n**int=x\n For example:\n is_simple_power(1, 4) => true\n is_simple_power(2, 2) => true\n is_simple_power(8, 2) => true\n is_simple_power(3, 2) => false\n is_simple_power(3, 1) => false\n is_simple_power(5, 3) => false\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(16, 2)== True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(143214, 16)== False, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(4, 2)==True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(9, 3)==True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(16, 4)==True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(24, 2)==False, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(128, 4)==False, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(12, 6)==False, \"This prints if this assert fails 1 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(1, 1)==True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate(1, 12)==True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "is_simple_power"} {"task": "code_en_b99c6cc2caf07badb3f000d4", "id": "b99c6cc2caf07badb3f000d4", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/120", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef maximum(arr, k):\n \"\"\"\n Given an array arr of integers and a positive integer k, return a sorted list \n of length k with the maximum k numbers in arr.\n\n Example 1:\n\n Input: arr = [-3, -4, 5], k = 3\n Output: [-4, -3, 5]\n\n Example 2:\n\n Input: arr = [4, -4, 4], k = 2\n Output: [4, 4]\n\n Example 3:\n\n Input: arr = [-3, 2, 1, 2, -1, -2, 1], k = 1\n Output: [2]\n\n Note:\n 1. The length of the array will be in the range of [1, 1000].\n 2. The elements in the array will be in the range of [-1000, 1000].\n 3. 0 <= k <= len(arr)\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns the k largest elements of arr in ascending sorted order, with duplicates preserved, e.g. [-4,-3,5] for ([-3,-4,5],3), [4,4] for ([4,-4,4],2) and [2] for Example 3. A correct alternative algorithm such as heapq.nlargest followed by a sort, or a sort followed by slicing, earns full credit."}, {"id": "c2", "description": "Handles k=0 by returning an empty list. The common sorted(arr)[-k:] pitfall is wrong here because it returns the whole array when k=0, and must be penalized as a wrong value on an input in the stated domain. Also handles k=len(arr), negative numbers and all-equal elements correctly."}, {"id": "c3", "description": "Output is a valid indented Python function body with no signature, docstring or Markdown fences, and it runs without syntax or indentation errors. It does not crash on any input in the stated domain."}, {"id": "c4", "description": "Uses an efficient approach, such as sorting or a heap at O(n log n) or better, with clear code. It does not mutate the input in a way that changes results, and adds no extra guards that alter outputs."}]}, "target_words": null, "reference": " if k == 0:\n return []\n arr.sort()\n ans = arr[-k:]\n return ans\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/120"}, "prefix": "\ndef maximum(arr, k):\n \"\"\"\n Given an array arr of integers and a positive integer k, return a sorted list \n of length k with the maximum k numbers in arr.\n\n Example 1:\n\n Input: arr = [-3, -4, 5], k = 3\n Output: [-4, -3, 5]\n\n Example 2:\n\n Input: arr = [4, -4, 4], k = 2\n Output: [4, 4]\n\n Example 3:\n\n Input: arr = [-3, 2, 1, 2, -1, -2, 1], k = 1\n Output: [2]\n\n Note:\n 1. The length of the array will be in the range of [1, 1000].\n 2. The elements in the array will be in the range of [-1000, 1000].\n 3. 0 <= k <= len(arr)\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([-3, -4, 5], 3) == [-4, -3, 5]\n assert candidate([4, -4, 4], 2) == [4, 4]\n assert candidate([-3, 2, 1, 2, -1, -2, 1], 1) == [2]\n assert candidate([123, -123, 20, 0 , 1, 2, -3], 3) == [2, 20, 123]\n assert candidate([-123, 20, 0 , 1, 2, -3], 4) == [0, 1, 2, 20]\n assert candidate([5, 15, 0, 3, -13, -8, 0], 7) == [-13, -8, 0, 0, 3, 5, 15]\n assert candidate([-1, 0, 2, 5, 3, -10], 2) == [3, 5]\n assert candidate([1, 0, 5, -7], 1) == [5]\n assert candidate([4, -4], 2) == [-4, 4]\n assert candidate([-10, 10], 2) == [-10, 10]\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([1, 2, 3, -23, 243, -400, 0], 0) == []\n\n", "entry_point": "maximum"} {"task": "code_en_bf8e9002ecac6d711006c985", "id": "bf8e9002ecac6d711006c985", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/107", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef even_odd_palindrome(n):\n \"\"\"\n Given a positive integer n, return a tuple that has the number of even and odd\n integer palindromes that fall within the range(1, n), inclusive.\n\n Example 1:\n\n Input: 3\n Output: (1, 2)\n Explanation:\n Integer palindrome are 1, 2, 3. one of them is even, and two of them are odd.\n\n Example 2:\n\n Input: 12\n Output: (4, 6)\n Explanation:\n Integer palindrome are 1, 2, 3, 4, 5, 6, 7, 8, 9, 11. four of them are even, and 6 of them are odd.\n\n Note:\n 1. 1 <= n <= 10^3\n 2. returned tuple has the number of even and odd integer palindromes respectively.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on inclusive range: counts palindromes in 1..n inclusive (n itself counted if palindrome), returning (even_count, odd_count) in that order; yields (1,2) for n=3 and (4,6) for n=12, and handles boundaries such as n=1 -> (0,1), n=9 -> (4,5), n=10 -> (4,5), n=11 -> (4,6), n=1000 -> (48,60)."}, {"id": "c2", "description": "Palindrome test and parity are correct: uses a valid check (e.g., str(i)==str(i)[::-1] or numeric reversal) that treats single digits as palindromes and rejects numbers like 10, 100; parity classified via i%2 on the number itself, not on digit count."}, {"id": "c3", "description": "Valid, executable function body: consistent indentation starting at function-body level, no repeated signature/docstring or Markdown, no syntax errors, returns a tuple, terminates for all 1<=n<=1000 without crashing; indentation that breaks the body lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: simple O(n·digits) or better approach appropriate for n<=10^3, clear variable names, no unnecessary guards that alter outputs or excessive commentary; any correct alternative algorithm gets full credit."}]}, "target_words": null, "reference": " def is_palindrome(n):\n return str(n) == str(n)[::-1]\n\n even_palindrome_count = 0\n odd_palindrome_count = 0\n\n for i in range(1, n+1):\n if i%2 == 1 and is_palindrome(i):\n odd_palindrome_count += 1\n elif i%2 == 0 and is_palindrome(i):\n even_palindrome_count += 1\n return (even_palindrome_count, odd_palindrome_count)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/107"}, "prefix": "\ndef even_odd_palindrome(n):\n \"\"\"\n Given a positive integer n, return a tuple that has the number of even and odd\n integer palindromes that fall within the range(1, n), inclusive.\n\n Example 1:\n\n Input: 3\n Output: (1, 2)\n Explanation:\n Integer palindrome are 1, 2, 3. one of them is even, and two of them are odd.\n\n Example 2:\n\n Input: 12\n Output: (4, 6)\n Explanation:\n Integer palindrome are 1, 2, 3, 4, 5, 6, 7, 8, 9, 11. four of them are even, and 6 of them are odd.\n\n Note:\n 1. 1 <= n <= 10^3\n 2. returned tuple has the number of even and odd integer palindromes respectively.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(123) == (8, 13)\n assert candidate(12) == (4, 6)\n assert candidate(3) == (1, 2)\n assert candidate(63) == (6, 8)\n assert candidate(25) == (5, 6)\n assert candidate(19) == (4, 6)\n assert candidate(9) == (4, 5), \"This prints if this assert fails 1 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(1) == (0, 1), \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "even_odd_palindrome"} {"task": "code_en_cb11722590367a3348c287d5", "id": "cb11722590367a3348c287d5", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/137", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef compare_one(a, b):\n \"\"\"\n Create a function that takes integers, floats, or strings representing\n real numbers, and returns the larger variable in its given variable type.\n Return None if the values are equal.\n Note: If a real number is represented as a string, the floating point might be . or ,\n\n compare_one(1, 2.5) ➞ 2.5\n compare_one(1, \"2,3\") ➞ \"2,3\"\n compare_one(\"5,1\", \"6\") ➞ \"6\"\n compare_one(\"1\", 1) ➞ None\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correct comparison: converts each argument to a numeric value for comparison (strings parsed after replacing ',' with '.'), returns the original argument unchanged in its original type (e.g. returns \"2,3\" not 2.3, 2.5 not '2.5'), and returns None when the numeric values are equal. All four docstring examples produce the stated outputs."}, {"id": "c2", "description": "Boundary cases handled: mixed int/float/string inputs such as (\"1\", 1) yield None, negative values such as (\"-1,5\", -2) compare correctly, zero comparisons work, and numeric inputs are never mutated or reformatted. No crash on valid inputs in the stated domain, for example non-string numbers must not have .replace called on them. A crash scores 6 or below."}, {"id": "c3", "description": "Valid output format: only an indented function body is returned, with no signature, docstring or Markdown fences. Indentation is consistent and the body is syntactically valid when placed under the given signature. If the first line's indentation breaks the body, this criterion is lowered."}, {"id": "c4", "description": "Algorithmic suitability and readability: a simple constant-time approach, with conversion done on copies and clear control flow. No extra guards that change outputs, such as rounding or returning floats, and no unnecessary complexity. Any correct alternative approach gets full credit."}]}, "target_words": null, "reference": " temp_a, temp_b = a, b\n if isinstance(temp_a, str): temp_a = temp_a.replace(',','.')\n if isinstance(temp_b, str): temp_b = temp_b.replace(',','.')\n if float(temp_a) == float(temp_b): return None\n return a if float(temp_a) > float(temp_b) else b \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/137"}, "prefix": "\ndef compare_one(a, b):\n \"\"\"\n Create a function that takes integers, floats, or strings representing\n real numbers, and returns the larger variable in its given variable type.\n Return None if the values are equal.\n Note: If a real number is represented as a string, the floating point might be . or ,\n\n compare_one(1, 2.5) ➞ 2.5\n compare_one(1, \"2,3\") ➞ \"2,3\"\n compare_one(\"5,1\", \"6\") ➞ \"6\"\n compare_one(\"1\", 1) ➞ None\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(1, 2) == 2\n assert candidate(1, 2.5) == 2.5\n assert candidate(2, 3) == 3\n assert candidate(5, 6) == 6\n assert candidate(1, \"2,3\") == \"2,3\"\n assert candidate(\"5,1\", \"6\") == \"6\"\n assert candidate(\"1\", \"2\") == \"2\"\n assert candidate(\"1\", 1) == None\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "compare_one"} {"task": "code_en_cb1e8baf9404a90853cc05b5", "id": "cb1e8baf9404a90853cc05b5", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/50", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef encode_shift(s: str):\n \"\"\"\n returns encoded string by shifting every character by 5 in the alphabet.\n \"\"\"\n return \"\".join([chr(((ord(ch) + 5 - ord(\"a\")) % 26) + ord(\"a\")) for ch in s])\n\n\ndef decode_shift(s: str):\n \"\"\"\n takes as input string encoded with encode_shift function. Returns decoded string.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correct inversion: the body returns a string that maps each character ch to chr(((ord(ch) - 5 - ord('a')) % 26) + ord('a')), or an equivalent, so that decode_shift(encode_shift(x)) == x for any lowercase string x. Mentally verify it on 'abc' -> encode 'fgh' -> decode 'abc', and on wraparound cases such as 'vwxyz' <-> 'abcde'. Wrong direction or missing modulo wraparound is a major failure."}, {"id": "c2", "description": "Boundary cases: the empty string returns '' without crashing. The modulo handling correctly wraps characters near 'a' (e.g., 'a','b','c','d','e' decode to 'v','w','x','y','z'). Because Python's % is non-negative, a negative intermediate is acceptable. Do not penalize the absence of guards for non-lowercase input, since the rubric's domain is output of encode_shift on lowercase text. Penalize extra guards that change outputs for valid encoded input, such as skipping characters."}, {"id": "c3", "description": "Syntactic validity: the response is only an indented function body with consistent indentation, starting at the function-body level, and it contains no signature, docstring repetition, or Markdown fences. The body must parse and execute as the body of decode_shift. Lower this score if the first line's indentation breaks the body."}, {"id": "c4", "description": "Algorithmic suitability and readability: the solution runs in linear time with a single pass, for example a join over a comprehension. It is concise and mirrors encode_shift clearly. Do not reward commentary or unnecessary complexity. A different correct approach, such as a translation table or calling encode_shift 21 times, receives full correctness credit; only gross inefficiency affects this criterion."}]}, "target_words": null, "reference": " return \"\".join([chr(((ord(ch) - 5 - ord(\"a\")) % 26) + ord(\"a\")) for ch in s])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/50"}, "prefix": "\n\ndef encode_shift(s: str):\n \"\"\"\n returns encoded string by shifting every character by 5 in the alphabet.\n \"\"\"\n return \"\".join([chr(((ord(ch) + 5 - ord(\"a\")) % 26) + ord(\"a\")) for ch in s])\n\n\ndef decode_shift(s: str):\n \"\"\"\n takes as input string encoded with encode_shift function. Returns decoded string.\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n from random import randint, choice\n import copy\n import string\n\n letters = string.ascii_lowercase\n for _ in range(100):\n str = ''.join(choice(letters) for i in range(randint(10, 20)))\n encoded_str = encode_shift(str)\n assert candidate(copy.deepcopy(encoded_str)) == str\n\n", "entry_point": "decode_shift"} {"task": "code_en_cc32bf774f0fb1cb03424c8f", "id": "cc32bf774f0fb1cb03424c8f", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/92", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef any_int(x, y, z):\n '''\n Create a function that takes 3 numbers.\n Returns true if one of the numbers is equal to the sum of the other two, and all numbers are integers.\n Returns false in any other cases.\n \n Examples\n any_int(5, 2, 7) ➞ True\n \n any_int(3, 2, 2) ➞ False\n\n any_int(3, -2, 1) ➞ True\n \n any_int(3.6, -2.2, 2) ➞ False\n \n\n \n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns True exactly when all three arguments are integers (isinstance int) and one equals the sum of the other two, checking all three permutations (x+y==z, x+z==y, y+z==x); produces True for (5,2,7) and (3,-2,1), False for (3,2,2) and (3.6,-2.2,2)."}, {"id": "c2", "description": "Rejects non-integer inputs even when the sum relation holds numerically, e.g. floats like (1.0, 2.0, 3.0) or (1.5, 1.5, 3) return False; the integer check must precede or gate the sum check rather than relying on value equality. Correctly handles zero and negative integers, e.g. (0,0,0) True and (-5,2,-3) True."}, {"id": "c3", "description": "The returned text is a syntactically valid, consistently indented function body (no signature, docstring, or Markdown fences) that runs without crashing on any numeric inputs and returns a boolean."}, {"id": "c4", "description": "Implementation is concise and readable with constant-time logic; no extraneous guards or commentary that alter outputs beyond the specified behavior."}]}, "target_words": null, "reference": " \n if isinstance(x,int) and isinstance(y,int) and isinstance(z,int):\n if (x+y==z) or (x+z==y) or (y+z==x):\n return True\n return False\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/92"}, "prefix": "\ndef any_int(x, y, z):\n '''\n Create a function that takes 3 numbers.\n Returns true if one of the numbers is equal to the sum of the other two, and all numbers are integers.\n Returns false in any other cases.\n \n Examples\n any_int(5, 2, 7) ➞ True\n \n any_int(3, 2, 2) ➞ False\n\n any_int(3, -2, 1) ➞ True\n \n any_int(3.6, -2.2, 2) ➞ False\n \n\n \n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(2, 3, 1)==True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(2.5, 2, 3)==False, \"This prints if this assert fails 2 (good for debugging!)\"\n assert candidate(1.5, 5, 3.5)==False, \"This prints if this assert fails 3 (good for debugging!)\"\n assert candidate(2, 6, 2)==False, \"This prints if this assert fails 4 (good for debugging!)\"\n assert candidate(4, 2, 2)==True, \"This prints if this assert fails 5 (good for debugging!)\"\n assert candidate(2.2, 2.2, 2.2)==False, \"This prints if this assert fails 6 (good for debugging!)\"\n assert candidate(-4, 6, 2)==True, \"This prints if this assert fails 7 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(2,1,1)==True, \"This prints if this assert fails 8 (also good for debugging!)\"\n assert candidate(3,4,7)==True, \"This prints if this assert fails 9 (also good for debugging!)\"\n assert candidate(3.0,4,7)==False, \"This prints if this assert fails 10 (also good for debugging!)\"\n\n", "entry_point": "any_int"} {"task": "code_en_cc71c18247d35c2dcf022fef", "id": "cc71c18247d35c2dcf022fef", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/135", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef can_arrange(arr):\n \"\"\"Create a function which returns the largest index of an element which\n is not greater than or equal to the element immediately preceding it. If\n no such element exists then return -1. The given array will not contain\n duplicate values.\n\n Examples:\n can_arrange([1,2,4,3,5]) = 3\n can_arrange([1,2,3]) = -1\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns the largest index i (i>=1) such that arr[i] < arr[i-1], scanning the whole array (not stopping at the first descent); e.g. [1,2,4,3,5] -> 3, [1,2,3] -> -1, [4,3,2,1] -> 3, [1,4,2,5,3] -> 4."}, {"id": "c2", "description": "Handles boundary inputs without crashing or index errors: empty list returns -1, single-element list returns -1, and index 0 is never returned or compared against arr[-1] (no wraparound comparison)."}, {"id": "c3", "description": "Returned body is syntactically valid Python with consistent indentation suitable for placement directly under the given signature/docstring, contains no Markdown or repeated signature, and actually returns an int in all paths."}, {"id": "c4", "description": "Uses a simple linear O(n) pass (forward tracking last descent or backward scan returning first found) with clear variable names; no unnecessary sorting, extra guards that alter outputs, or excessive complexity."}]}, "target_words": null, "reference": " ind=-1\n i=1\n while i 12\n solution([3, 3, 3, 3, 3]) ==> 9\n solution([30, 13, 24, 321]) ==>0\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: returns the sum of elements at even indices (0, 2, 4, ...) that are odd. It yields 12 for [5,8,7,1], 9 for [3,3,3,3,3] and 0 for [30,13,24,321]. Position parity uses 0-based indices, and value parity is checked separately."}, {"id": "c2", "description": "Edge cases: negative odd integers count as odd. In Python, x % 2 == 1 is true for negatives, so the check must not use a condition like x % 2 == 1 only for positives. The function returns 0 when no qualifying elements exist and handles a single-element list. It does not crash or loop forever on any non-empty integer list. Zero counts as even."}, {"id": "c3", "description": "Validity: the output is only an indented function body, with no signature, docstring or Markdown fences. Indentation is consistent, starting from the first line, so the body parses as valid Python and returns an int."}, {"id": "c4", "description": "Algorithm and readability: uses a single linear pass, such as a generator over enumerate or slicing lst[::2], with no unnecessary complexity. Extra guards or commentary must not change the output."}]}, "target_words": null, "reference": " return sum([x for idx, x in enumerate(lst) if idx%2==0 and x%2==1])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/121"}, "prefix": "\ndef solution(lst):\n \"\"\"Given a non-empty list of integers, return the sum of all of the odd elements that are in even positions.\n \n\n Examples\n solution([5, 8, 7, 1]) ==> 12\n solution([3, 3, 3, 3, 3]) ==> 9\n solution([30, 13, 24, 321]) ==>0\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([5, 8, 7, 1]) == 12\n assert candidate([3, 3, 3, 3, 3]) == 9\n assert candidate([30, 13, 24, 321]) == 0\n assert candidate([5, 9]) == 5\n assert candidate([2, 4, 8]) == 0\n assert candidate([30, 13, 23, 32]) == 23\n assert candidate([3, 13, 2, 9]) == 3\n\n # Check some edge cases that are easy to work out by hand.\n\n", "entry_point": "solution"} {"task": "code_en_d23be443ebef52118aaaaa1b", "id": "d23be443ebef52118aaaaa1b", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/114", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef minSubArraySum(nums):\n \"\"\"\n Given an array of integers nums, find the minimum sum of any non-empty sub-array\n of nums.\n Example\n minSubArraySum([2, 3, 4, 1, 2, 4]) == 1\n minSubArraySum([-1, -2, -3]) == -6\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core cases: returns the minimum sum over all non-empty contiguous sub-arrays, e.g. 1 for [2,3,4,1,2,4] and -6 for [-1,-2,-3]. A mixed case such as [2,-1,-2,3,-4] gives -4 and [1,-1,-1,1] gives -2. A Kadane-style minimum or a brute-force enumeration are equally acceptable."}, {"id": "c2", "description": "Boundary cases: an all-positive input returns its smallest element, not 0, because the empty sub-array is excluded. A single-element list returns that element. Zeros, large magnitudes and duplicate values are handled. An implementation that initializes the running minimum to 0, and so returns 0 for all-positive input, is a moderate-to-severe correctness flaw. Empty input is outside the stated domain, so any reasonable behavior is acceptable."}, {"id": "c3", "description": "Validity and readability: the returned text is only a correctly indented function body with no signature, docstring or Markdown. It is syntactically valid and terminates. A crash or non-termination on valid input scores 6 or below. Inconsistent first-line indentation that breaks the body lowers this criterion. The code is clear and does not add output-changing guards."}, {"id": "c4", "description": "Algorithmic suitability: O(n) or O(n^2) approaches that are correct receive full credit. Exponential approaches, or approaches impractical for large lists, are penalized modestly. Extra complexity is not rewarded."}]}, "target_words": null, "reference": " max_sum = 0\n s = 0\n for num in nums:\n s += -num\n if (s < 0):\n s = 0\n max_sum = max(s, max_sum)\n if max_sum == 0:\n max_sum = max(-i for i in nums)\n min_sum = -max_sum\n return min_sum\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/114"}, "prefix": "\ndef minSubArraySum(nums):\n \"\"\"\n Given an array of integers nums, find the minimum sum of any non-empty sub-array\n of nums.\n Example\n minSubArraySum([2, 3, 4, 1, 2, 4]) == 1\n minSubArraySum([-1, -2, -3]) == -6\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([2, 3, 4, 1, 2, 4]) == 1, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([-1, -2, -3]) == -6\n assert candidate([-1, -2, -3, 2, -10]) == -14\n assert candidate([-9999999999999999]) == -9999999999999999\n assert candidate([0, 10, 20, 1000000]) == 0\n assert candidate([-1, -2, -3, 10, -5]) == -6\n assert candidate([100, -1, -2, -3, 10, -5]) == -6\n assert candidate([10, 11, 13, 8, 3, 4]) == 3\n assert candidate([100, -33, 32, -1, 0, -2]) == -33\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([-10]) == -10, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate([7]) == 7\n assert candidate([1, -1]) == -1\n", "entry_point": "minSubArraySum"} {"task": "code_en_da44faff5d21601dbec8ce65", "id": "da44faff5d21601dbec8ce65", "domain": "code", "language": "en", "split": "train", "group": "HumanEval/31", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef is_prime(n):\n \"\"\"Return true if a given number is prime, and false otherwise.\n >>> is_prime(6)\n False\n >>> is_prime(101)\n True\n >>> is_prime(11)\n True\n >>> is_prime(13441)\n True\n >>> is_prime(61)\n True\n >>> is_prime(4)\n False\n >>> is_prime(1)\n False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns correct booleans on all docstring examples (6→False, 101→True, 11→True, 13441→True, 61→True, 4→False, 1→False) and on adversarial inputs: 0, negative integers, 1 return False; 2 and 3 return True; even composites and squares of primes (e.g. 9, 25, 49) return False. Crashes or non-termination on any integer input score 6 or below."}, {"id": "c2", "description": "Uses an algorithmically suitable primality test, such as trial division up to sqrt(n) with a correct loop bound that includes the square root itself. Full credit goes to any correct alternative. A loop bound off by one that misclassifies perfect squares of primes is penalized according to how plausible those inputs are."}, {"id": "c3", "description": "The returned body is syntactically valid Python, consistently indented as a function body, without a repeated signature, docstring or Markdown fences. Indentation errors on the first line or elsewhere that break execution lower this criterion substantially."}, {"id": "c4", "description": "Code is readable and concise, with clear control flow and no extraneous guards or commentary that alter outputs. Style is weighted far less than correctness."}]}, "target_words": null, "reference": " if n < 2:\n return False\n for k in range(2, n - 1):\n if n % k == 0:\n return False\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/31"}, "prefix": "\n\ndef is_prime(n):\n \"\"\"Return true if a given number is prime, and false otherwise.\n >>> is_prime(6)\n False\n >>> is_prime(101)\n True\n >>> is_prime(11)\n True\n >>> is_prime(13441)\n True\n >>> is_prime(61)\n True\n >>> is_prime(4)\n False\n >>> is_prime(1)\n False\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(6) == False\n assert candidate(101) == True\n assert candidate(11) == True\n assert candidate(13441) == True\n assert candidate(61) == True\n assert candidate(4) == False\n assert candidate(1) == False\n assert candidate(5) == True\n assert candidate(11) == True\n assert candidate(17) == True\n assert candidate(5 * 17) == False\n assert candidate(11 * 7) == False\n assert candidate(13441 * 19) == False\n\n", "entry_point": "is_prime"} {"task": "code_en_db1811f45910d4332a2e401c", "id": "db1811f45910d4332a2e401c", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/74", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef total_match(lst1, lst2):\n '''\n Write a function that accepts two lists of strings and returns the list that has \n total number of chars in the all strings of the list less than the other list.\n\n if the two lists have the same number of chars, return the first list.\n\n Examples\n total_match([], []) ➞ []\n total_match(['hi', 'admin'], ['hI', 'Hi']) ➞ ['hI', 'Hi']\n total_match(['hi', 'admin'], ['hi', 'hi', 'admin', 'project']) ➞ ['hi', 'admin']\n total_match(['hi', 'admin'], ['hI', 'hi', 'hi']) ➞ ['hI', 'hi', 'hi']\n total_match(['4'], ['1', '2', '3', '4', '5']) ➞ ['4']\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: The body computes the total character count of each list (the sum of len over its strings) and returns lst1 when its total is less than or equal to lst2's total, otherwise lst2. It returns the original list object, not a copy or a sorted or modified version. Tracing all five docstring examples gives the expected outputs."}, {"id": "c2", "description": "Boundary cases: Ties, including the case where both lists are empty, return lst1. One empty list against a non-empty list returns the empty one. Lists containing empty strings are counted correctly with zero chars. A wrong tie-break, such as using < where <= is needed, is a serious error because the docstring names that case. A crash on empty input scores 6 or below."}, {"id": "c3", "description": "Validity and suitability: The returned text is only a validly indented function body with no signature, docstring or Markdown. It runs without syntax or indentation errors. It uses a linear pass, for example sum(len(s) for s in lst), and does not count list length instead of characters."}, {"id": "c4", "description": "Readability: The logic is concise and clear. It has no extraneous guards or commentary that change outputs, and no unnecessary complexity."}]}, "target_words": null, "reference": " l1 = 0\n for st in lst1:\n l1 += len(st)\n \n l2 = 0\n for st in lst2:\n l2 += len(st)\n \n if l1 <= l2:\n return lst1\n else:\n return lst2\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/74"}, "prefix": "\ndef total_match(lst1, lst2):\n '''\n Write a function that accepts two lists of strings and returns the list that has \n total number of chars in the all strings of the list less than the other list.\n\n if the two lists have the same number of chars, return the first list.\n\n Examples\n total_match([], []) ➞ []\n total_match(['hi', 'admin'], ['hI', 'Hi']) ➞ ['hI', 'Hi']\n total_match(['hi', 'admin'], ['hi', 'hi', 'admin', 'project']) ➞ ['hi', 'admin']\n total_match(['hi', 'admin'], ['hI', 'hi', 'hi']) ➞ ['hI', 'hi', 'hi']\n total_match(['4'], ['1', '2', '3', '4', '5']) ➞ ['4']\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([], []) == []\n assert candidate(['hi', 'admin'], ['hi', 'hi']) == ['hi', 'hi']\n assert candidate(['hi', 'admin'], ['hi', 'hi', 'admin', 'project']) == ['hi', 'admin']\n assert candidate(['4'], ['1', '2', '3', '4', '5']) == ['4']\n assert candidate(['hi', 'admin'], ['hI', 'Hi']) == ['hI', 'Hi']\n assert candidate(['hi', 'admin'], ['hI', 'hi', 'hi']) == ['hI', 'hi', 'hi']\n assert candidate(['hi', 'admin'], ['hI', 'hi', 'hii']) == ['hi', 'admin']\n\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate([], ['this']) == []\n assert candidate(['this'], []) == []\n\n", "entry_point": "total_match"} {"task": "code_en_dd19baf8b8f9b5bf042565a4", "id": "dd19baf8b8f9b5bf042565a4", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/158", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef find_max(words):\n \"\"\"Write a function that accepts a list of strings.\n The list contains different words. Return the word with maximum number\n of unique characters. If multiple strings have maximum number of unique\n characters, return the one which comes first in lexicographical order.\n\n find_max([\"name\", \"of\", \"string\"]) == \"string\"\n find_max([\"name\", \"enam\", \"game\"]) == \"enam\"\n find_max([\"aaaaaaa\", \"bb\" ,\"cc\"]) == \"\"aaaaaaa\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns the word with the maximum count of distinct characters (len(set(word))), and breaks ties by choosing the lexicographically smallest word, e.g. via sorted(words, key=lambda w: (-len(set(w)), w))[0] or an equivalent comparison. It must produce \"string\", \"enam\" and \"aaaaaaa\" on the three docstring examples."}, {"id": "c2", "description": "The tie-break is lexicographic order, not first occurrence in the list. For example, [\"name\",\"enam\",\"game\"] must yield \"enam\", and [\"bb\",\"cc\"] must yield \"bb\". Words that are anagrams or that repeat characters are scored by unique characters, not by total length."}, {"id": "c3", "description": "Handles boundary inputs without crashing or looping forever. A single-element list returns that element. Words that include the empty string, or a list where all words are equal, are handled. Large lists are processed in roughly O(total characters) or O(n log n) time. Behavior on an empty list is not specified, so do not penalize any reasonable choice for it."}, {"id": "c4", "description": "The returned body is syntactically valid Python with consistent indentation suitable for the function body. It contains no repeated signature or docstring and no Markdown fences, and it is reasonably readable."}]}, "target_words": null, "reference": " return sorted(words, key = lambda x: (-len(set(x)), x))[0]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/158"}, "prefix": "\ndef find_max(words):\n \"\"\"Write a function that accepts a list of strings.\n The list contains different words. Return the word with maximum number\n of unique characters. If multiple strings have maximum number of unique\n characters, return the one which comes first in lexicographical order.\n\n find_max([\"name\", \"of\", \"string\"]) == \"string\"\n find_max([\"name\", \"enam\", \"game\"]) == \"enam\"\n find_max([\"aaaaaaa\", \"bb\" ,\"cc\"]) == \"\"aaaaaaa\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert (candidate([\"name\", \"of\", \"string\"]) == \"string\"), \"t1\"\n assert (candidate([\"name\", \"enam\", \"game\"]) == \"enam\"), 't2'\n assert (candidate([\"aaaaaaa\", \"bb\", \"cc\"]) == \"aaaaaaa\"), 't3'\n assert (candidate([\"abc\", \"cba\"]) == \"abc\"), 't4'\n assert (candidate([\"play\", \"this\", \"game\", \"of\",\"footbott\"]) == \"footbott\"), 't5'\n assert (candidate([\"we\", \"are\", \"gonna\", \"rock\"]) == \"gonna\"), 't6'\n assert (candidate([\"we\", \"are\", \"a\", \"mad\", \"nation\"]) == \"nation\"), 't7'\n assert (candidate([\"this\", \"is\", \"a\", \"prrk\"]) == \"this\"), 't8'\n\n # Check some edge cases that are easy to work out by hand.\n assert (candidate([\"b\"]) == \"b\"), 't9'\n assert (candidate([\"play\", \"play\", \"play\"]) == \"play\"), 't10'\n\n", "entry_point": "find_max"} {"task": "code_en_e0bbd9622e73f6b8ea9d89c6", "id": "e0bbd9622e73f6b8ea9d89c6", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/100", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef make_a_pile(n):\n \"\"\"\n Given a positive integer n, you have to make a pile of n levels of stones.\n The first level has n stones.\n The number of stones in the next level is:\n - the next odd number if n is odd.\n - the next even number if n is even.\n Return the number of stones in each level in a list, where element at index\n i represents the number of stones in the level (i+1).\n\n Examples:\n >>> make_a_pile(3)\n [3, 5, 7]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: returns a list of exactly n integers where element i equals n + 2*i (e.g., make_a_pile(3) == [3, 5, 7], make_a_pile(4) == [4, 6, 8, 10], make_a_pile(1) == [1]). Wrong parity handling, off-by-one length, or wrong step size is a major failure."}, {"id": "c2", "description": "Boundary handling within the stated domain (positive integers): n=1 returns [1], and large n (e.g., 10**5) completes without crash or non-termination in linear time. Crash or infinite loop on any positive n scores 6 or below. No extra guards that alter outputs for valid inputs."}, {"id": "c3", "description": "Validity of the returned body: it is only the indented function body (no signature, docstring, or Markdown fences), with consistent indentation that parses correctly when placed under the given signature, and it returns a list rather than printing or yielding."}, {"id": "c4", "description": "Algorithmic suitability and readability: uses a simple O(n) construction (such as a list comprehension or range with step 2) that is clear. Any correct alternative earns full credit; unnecessary complexity or commentary does not."}]}, "target_words": null, "reference": " return [n + 2*i for i in range(n)]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/100"}, "prefix": "\ndef make_a_pile(n):\n \"\"\"\n Given a positive integer n, you have to make a pile of n levels of stones.\n The first level has n stones.\n The number of stones in the next level is:\n - the next odd number if n is odd.\n - the next even number if n is even.\n Return the number of stones in each level in a list, where element at index\n i represents the number of stones in the level (i+1).\n\n Examples:\n >>> make_a_pile(3)\n [3, 5, 7]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(3) == [3, 5, 7], \"Test 3\"\n assert candidate(4) == [4,6,8,10], \"Test 4\"\n assert candidate(5) == [5, 7, 9, 11, 13]\n assert candidate(6) == [6, 8, 10, 12, 14, 16]\n assert candidate(8) == [8, 10, 12, 14, 16, 18, 20, 22]\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "make_a_pile"} {"task": "code_en_e27e86e99896fd5a130ca8db", "id": "e27e86e99896fd5a130ca8db", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/113", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef odd_count(lst):\n \"\"\"Given a list of strings, where each string consists of only digits, return a list.\n Each element i of the output should be \"the number of odd elements in the\n string i of the input.\" where all the i's should be replaced by the number\n of odd digits in the i'th string of the input.\n\n >>> odd_count(['1234567'])\n [\"the number of odd elements 4n the str4ng 4 of the 4nput.\"]\n >>> odd_count(['3',\"11111111\"])\n [\"the number of odd elements 1n the str1ng 1 of the 1nput.\",\n \"the number of odd elements 8n the str8ng 8 of the 8nput.\"]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Output strings exactly match the template \"the number of odd elements Nn the strNng N of the Nnput.\" with all four 'i' placeholders replaced by the count N, and only those four. Other letters such as the 'i' in 'elements' are unaffected. Both docstring examples are reproduced exactly: ['1234567'] gives N=4, and ['3','11111111'] gives N=1 and 8."}, {"id": "c2", "description": "N is the count of odd digits (1,3,5,7,9) in each string. It is computed correctly via int(d)%2==1 or equivalent and is not confused with the count of odd positions or with string length. One output element is produced per input string, in input order."}, {"id": "c3", "description": "Boundary cases are handled without crashing. An empty list returns []. A string with no odd digits (e.g. '2468' or '') yields N=0 in all four places. Multi-digit counts (e.g. 12 odd digits) are inserted as the full number '12'."}, {"id": "c4", "description": "The returned body is syntactically valid, consistently indented as a function body, and contains only the body without the signature or Markdown. It is linear in total input size and reasonably readable."}]}, "target_words": null, "reference": " res = []\n for arr in lst:\n n = sum(int(d)%2==1 for d in arr)\n res.append(\"the number of odd elements \" + str(n) + \"n the str\"+ str(n) +\"ng \"+ str(n) +\" of the \"+ str(n) +\"nput.\")\n return res\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/113"}, "prefix": "\ndef odd_count(lst):\n \"\"\"Given a list of strings, where each string consists of only digits, return a list.\n Each element i of the output should be \"the number of odd elements in the\n string i of the input.\" where all the i's should be replaced by the number\n of odd digits in the i'th string of the input.\n\n >>> odd_count(['1234567'])\n [\"the number of odd elements 4n the str4ng 4 of the 4nput.\"]\n >>> odd_count(['3',\"11111111\"])\n [\"the number of odd elements 1n the str1ng 1 of the 1nput.\",\n \"the number of odd elements 8n the str8ng 8 of the 8nput.\"]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(['1234567']) == [\"the number of odd elements 4n the str4ng 4 of the 4nput.\"], \"Test 1\"\n assert candidate(['3',\"11111111\"]) == [\"the number of odd elements 1n the str1ng 1 of the 1nput.\", \"the number of odd elements 8n the str8ng 8 of the 8nput.\"], \"Test 2\"\n assert candidate(['271', '137', '314']) == [\n 'the number of odd elements 2n the str2ng 2 of the 2nput.',\n 'the number of odd elements 3n the str3ng 3 of the 3nput.',\n 'the number of odd elements 2n the str2ng 2 of the 2nput.'\n ]\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "odd_count"} {"task": "code_en_e6719b744b1907881a8dcaa9", "id": "e6719b744b1907881a8dcaa9", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/10", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\n\ndef is_palindrome(string: str) -> bool:\n \"\"\" Test if given string is a palindrome \"\"\"\n return string == string[::-1]\n\n\ndef make_palindrome(string: str) -> str:\n \"\"\" Find the shortest palindrome that begins with a supplied string.\n Algorithm idea is simple:\n - Find the longest postfix of supplied string that is a palindrome.\n - Append to the end of the string reverse of a string prefix that comes before the palindromic suffix.\n >>> make_palindrome('')\n ''\n >>> make_palindrome('cat')\n 'catac'\n >>> make_palindrome('cata')\n 'catac'\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields '' for '', 'catac' for 'cat', and 'catac' for 'cata'. The result always starts with the input string, is a palindrome, and is the shortest such string. It does this by appending the reverse of the prefix that precedes the longest palindromic suffix. Any algorithm meeting this definition earns full credit."}, {"id": "c2", "description": "Boundary cases: the body handles the empty string without crashing or looping and returns ''. It returns a single character unchanged and returns an already-palindromic input such as 'aba' or 'aa' unchanged. Inputs like 'ab' -> 'aba' and 'abc' -> 'abcba' are correct, and the code introduces no off-by-one error in the suffix search. Non-termination or an exception on any string input scores 6 or below."}, {"id": "c3", "description": "Validity and form: the output is only an indented function body with no signature, docstring repetition, or Markdown. The indentation is consistent and syntactically valid as given. The body may call the provided is_palindrome helper. A first line whose indentation breaks the body lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is a reasonable O(n^2) suffix scan, or a better one such as KMP, and is clearly structured. No extraneous guards alter outputs, and the code avoids needless complexity. Style is weighed only after correctness."}]}, "target_words": null, "reference": " if not string:\n return ''\n\n beginning_of_suffix = 0\n\n while not is_palindrome(string[beginning_of_suffix:]):\n beginning_of_suffix += 1\n\n return string + string[:beginning_of_suffix][::-1]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/10"}, "prefix": "\n\ndef is_palindrome(string: str) -> bool:\n \"\"\" Test if given string is a palindrome \"\"\"\n return string == string[::-1]\n\n\ndef make_palindrome(string: str) -> str:\n \"\"\" Find the shortest palindrome that begins with a supplied string.\n Algorithm idea is simple:\n - Find the longest postfix of supplied string that is a palindrome.\n - Append to the end of the string reverse of a string prefix that comes before the palindromic suffix.\n >>> make_palindrome('')\n ''\n >>> make_palindrome('cat')\n 'catac'\n >>> make_palindrome('cata')\n 'catac'\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('') == ''\n assert candidate('x') == 'x'\n assert candidate('xyz') == 'xyzyx'\n assert candidate('xyx') == 'xyx'\n assert candidate('jerry') == 'jerryrrej'\n", "entry_point": "make_palindrome"} {"task": "code_en_ee0d94dec8329601eb642013", "id": "ee0d94dec8329601eb642013", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/152", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef compare(game,guess):\n \"\"\"I think we all remember that feeling when the result of some long-awaited\n event is finally known. The feelings and thoughts you have at that moment are\n definitely worth noting down and comparing.\n Your task is to determine if a person correctly guessed the results of a number of matches.\n You are given two arrays of scores and guesses of equal length, where each index shows a match. \n Return an array of the same length denoting how far off each guess was. If they have guessed correctly,\n the value is 0, and if not, the value is the absolute difference between the guess and the score.\n \n \n example:\n\n compare([1,2,3,4,5,1],[1,2,3,4,2,-2]) -> [0,0,0,0,3,3]\n compare([0,5,0,0,0,4],[4,1,1,0,0,-2]) -> [4,4,1,0,0,6]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Body returns a list where each element equals abs(game[i] - guess[i]) for the corresponding index, producing [0,0,0,0,3,3] and [4,4,1,0,0,6] on the docstring examples; correct guesses yield 0."}, {"id": "c2", "description": "Handles boundary cases correctly: negative scores or guesses (absolute difference, not signed), zeros, empty input lists (returns []), and large integers, without crashing or mutating the inputs."}, {"id": "c3", "description": "Returned body is syntactically valid and consistently indented as a function body (first line indentation does not break it), with no repeated signature, docstring, or Markdown fences."}, {"id": "c4", "description": "Uses a simple linear, readable approach (e.g., list comprehension with zip or index iteration) without unnecessary guards that alter outputs or extraneous complexity."}]}, "target_words": null, "reference": " return [abs(x-y) for x,y in zip(game,guess)]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/152"}, "prefix": "\ndef compare(game,guess):\n \"\"\"I think we all remember that feeling when the result of some long-awaited\n event is finally known. The feelings and thoughts you have at that moment are\n definitely worth noting down and comparing.\n Your task is to determine if a person correctly guessed the results of a number of matches.\n You are given two arrays of scores and guesses of equal length, where each index shows a match. \n Return an array of the same length denoting how far off each guess was. If they have guessed correctly,\n the value is 0, and if not, the value is the absolute difference between the guess and the score.\n \n \n example:\n\n compare([1,2,3,4,5,1],[1,2,3,4,2,-2]) -> [0,0,0,0,3,3]\n compare([0,5,0,0,0,4],[4,1,1,0,0,-2]) -> [4,4,1,0,0,6]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([1,2,3,4,5,1],[1,2,3,4,2,-2])==[0,0,0,0,3,3], \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([0,0,0,0,0,0],[0,0,0,0,0,0])==[0,0,0,0,0,0], \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([1,2,3],[-1,-2,-3])==[2,4,6], \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate([1,2,3,5],[-1,2,3,4])==[2,0,0,1], \"This prints if this assert fails 1 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "compare"} {"task": "code_en_f5966842cf9fff1c79eba6ea", "id": "f5966842cf9fff1c79eba6ea", "domain": "code", "language": "en", "split": "test", "group": "HumanEval/70", "prompt": "Implement the Python function specified below. Return ONLY the indented function body, without repeating the signature/docstring or using Markdown. Correctness and edge cases matter more than style.\n\n\ndef strange_sort_list(lst):\n '''\n Given list of integers, return list in strange order.\n Strange sorting, is when you start with the minimum value,\n then maximum of the remaining integers, then minimum and so on.\n\n Examples:\n strange_sort_list([1, 2, 3, 4]) == [1, 4, 2, 3]\n strange_sort_list([5, 5, 5, 5]) == [5, 5, 5, 5]\n strange_sort_list([]) == []\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Functional correctness: the body returns a new list that alternates min of remaining, then max of remaining, until all elements are used. It produces [1,4,2,3] for [1,2,3,4] and [5,5,5,5] for [5,5,5,5]. It also handles odd lengths, e.g. [1,2,3,4,5] -> [1,5,2,4,3], and unsorted input, e.g. [5,6,7,8,9] -> [5,9,6,8,7]. Any correct algorithm earns full credit, whether it sorts and uses two pointers or repeatedly removes min/max."}, {"id": "c2", "description": "Boundary cases: the body handles the empty list (returns []), a single element, duplicates mixed with distinct values (e.g. [0,2,2,2,5,5,-5,-5] -> [-5,5,-5,5,0,2,2,2]), negative numbers and an all-equal list. It must not crash or loop forever, for example by popping from an empty list or by a loop condition that never ends. A crash or hang on any list of integers caps this criterion at 6 or below."}, {"id": "c3", "description": "Syntactic validity and format: the output is only an indented function body with consistent indentation, starting from the first line. It contains no signature, no docstring repetition and no Markdown fences, and it parses as valid Python when placed under the given signature. If it mutates the caller's list, that counts as a minor flaw only if outputs are unaffected."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is reasonable, e.g. O(n log n) sort with two pointers, or O(n^2) repeated min/max, which is acceptable. The code is clear and concise. It has no extra guards or special cases that change outputs, and it gets no credit for added commentary or complexity."}]}, "target_words": null, "reference": " res, switch = [], True\n while lst:\n res.append(min(lst) if switch else max(lst))\n lst.remove(res[-1])\n switch = not switch\n return res\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/70"}, "prefix": "\ndef strange_sort_list(lst):\n '''\n Given list of integers, return list in strange order.\n Strange sorting, is when you start with the minimum value,\n then maximum of the remaining integers, then minimum and so on.\n\n Examples:\n strange_sort_list([1, 2, 3, 4]) == [1, 4, 2, 3]\n strange_sort_list([5, 5, 5, 5]) == [5, 5, 5, 5]\n strange_sort_list([]) == []\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([1, 2, 3, 4]) == [1, 4, 2, 3]\n assert candidate([5, 6, 7, 8, 9]) == [5, 9, 6, 8, 7]\n assert candidate([1, 2, 3, 4, 5]) == [1, 5, 2, 4, 3]\n assert candidate([5, 6, 7, 8, 9, 1]) == [1, 9, 5, 8, 6, 7]\n assert candidate([5, 5, 5, 5]) == [5, 5, 5, 5]\n assert candidate([]) == []\n assert candidate([1,2,3,4,5,6,7,8]) == [1, 8, 2, 7, 3, 6, 4, 5]\n assert candidate([0,2,2,2,5,5,-5,-5]) == [-5, 5, -5, 5, 0, 2, 2, 2]\n assert candidate([111111]) == [111111]\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "strange_sort_list"} {"task": "code_ru_02ab772a7e06536a92ac8d22", "id": "02ab772a7e06536a92ac8d22", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/129", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef minPath(grid, k):\n \"\"\"\n Given a grid with N rows and N columns (N >= 2) and a positive integer k, \n each cell of the grid contains a value. Every integer in the range [1, N * N]\n inclusive appears exactly once on the cells of the grid.\n\n You have to find the minimum path of length k in the grid. You can start\n from any cell, and in each step you can move to any of the neighbor cells,\n in other words, you can go to cells which share an edge with you current\n cell.\n Please note that a path of length k means visiting exactly k cells (not\n necessarily distinct).\n You CANNOT go off the grid.\n A path A (of length k) is considered less than a path B (of length k) if\n after making the ordered lists of the values on the cells that A and B go\n through (let's call them lst_A and lst_B), lst_A is lexicographically less\n than lst_B, in other words, there exist an integer index i (1 <= i <= k)\n such that lst_A[i] < lst_B[i] and for any j (1 <= j < i) we have\n lst_A[j] = lst_B[j].\n It is guaranteed that the answer is unique.\n Return an ordered list of the values on the cells that the minimum path go through.\n\n Examples:\n\n Input: grid = [ [1,2,3], [4,5,6], [7,8,9]], k = 3\n Output: [1, 2, 1]\n\n Input: grid = [ [5,9,3], [4,1,6], [7,8,2]], k = 1\n Output: [1]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of path: the body returns a list of length k that alternates starting with 1 and the smallest neighbor of the cell containing 1 (e.g. [1,2,1] for the first example, [1] for k=1). Mentally execute on both docstring examples and on cases where 1 sits in a corner, on an edge, or in the interior, and for even and odd k. Wrong values on these common inputs are major failures."}, {"id": "c2", "description": "Boundary handling: the neighbor lookup never indexes off the grid. It must not use negative indices that silently wrap around in Python. It must handle N=2, k=1 and large k without crashing or looping forever. A crash or non-termination on valid input (N>=2, k>=1) caps this criterion at 6 or below."}, {"id": "c3", "description": "Algorithmic suitability: the approach runs in roughly O(N^2 + k) time, for example by locating 1 and alternating it with its minimum neighbor. Any other correct algorithm gets full credit. Exhaustive path enumeration that is exponential in k should be penalized, because it is impractical for large k."}, {"id": "c4", "description": "Validity and readability: the returned text is only the function body, indented correctly for placement under the signature. It contains no signature, no docstring and no Markdown fences, and it is syntactically valid Python. Variable names and structure are clear. Lower this score if first-line indentation or extra wrapper text breaks the body."}]}, "target_words": null, "reference": " n = len(grid)\n val = n * n + 1\n for i in range(n):\n for j in range(n):\n if grid[i][j] == 1:\n temp = []\n if i != 0:\n temp.append(grid[i - 1][j])\n\n if j != 0:\n temp.append(grid[i][j - 1])\n\n if i != n - 1:\n temp.append(grid[i + 1][j])\n\n if j != n - 1:\n temp.append(grid[i][j + 1])\n\n val = min(temp)\n\n ans = []\n for i in range(k):\n if i % 2 == 0:\n ans.append(1)\n else:\n ans.append(val)\n return ans\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/129", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef minPath(grid, k):\n \"\"\"\n Given a grid with N rows and N columns (N >= 2) and a positive integer k, \n each cell of the grid contains a value. Every integer in the range [1, N * N]\n inclusive appears exactly once on the cells of the grid.\n\n You have to find the minimum path of length k in the grid. You can start\n from any cell, and in each step you can move to any of the neighbor cells,\n in other words, you can go to cells which share an edge with you current\n cell.\n Please note that a path of length k means visiting exactly k cells (not\n necessarily distinct).\n You CANNOT go off the grid.\n A path A (of length k) is considered less than a path B (of length k) if\n after making the ordered lists of the values on the cells that A and B go\n through (let's call them lst_A and lst_B), lst_A is lexicographically less\n than lst_B, in other words, there exist an integer index i (1 <= i <= k)\n such that lst_A[i] < lst_B[i] and for any j (1 <= j < i) we have\n lst_A[j] = lst_B[j].\n It is guaranteed that the answer is unique.\n Return an ordered list of the values on the cells that the minimum path go through.\n\n Examples:\n\n Input: grid = [ [1,2,3], [4,5,6], [7,8,9]], k = 3\n Output: [1, 2, 1]\n\n Input: grid = [ [5,9,3], [4,1,6], [7,8,2]], k = 1\n Output: [1]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n print\n assert candidate([[1, 2, 3], [4, 5, 6], [7, 8, 9]], 3) == [1, 2, 1]\n assert candidate([[5, 9, 3], [4, 1, 6], [7, 8, 2]], 1) == [1]\n assert candidate([[1, 2, 3, 4], [5, 6, 7, 8], [9, 10, 11, 12], [13, 14, 15, 16]], 4) == [1, 2, 1, 2]\n assert candidate([[6, 4, 13, 10], [5, 7, 12, 1], [3, 16, 11, 15], [8, 14, 9, 2]], 7) == [1, 10, 1, 10, 1, 10, 1]\n assert candidate([[8, 14, 9, 2], [6, 4, 13, 15], [5, 7, 1, 12], [3, 10, 11, 16]], 5) == [1, 7, 1, 7, 1]\n assert candidate([[11, 8, 7, 2], [5, 16, 14, 4], [9, 3, 15, 6], [12, 13, 10, 1]], 9) == [1, 6, 1, 6, 1, 6, 1, 6, 1]\n assert candidate([[12, 13, 10, 1], [9, 3, 15, 6], [5, 16, 14, 4], [11, 8, 7, 2]], 12) == [1, 6, 1, 6, 1, 6, 1, 6, 1, 6, 1, 6]\n assert candidate([[2, 7, 4], [3, 1, 5], [6, 8, 9]], 8) == [1, 3, 1, 3, 1, 3, 1, 3]\n assert candidate([[6, 1, 5], [3, 8, 9], [2, 7, 4]], 8) == [1, 5, 1, 5, 1, 5, 1, 5]\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([[1, 2], [3, 4]], 10) == [1, 2, 1, 2, 1, 2, 1, 2, 1, 2]\n assert candidate([[1, 3], [3, 2]], 10) == [1, 3, 1, 3, 1, 3, 1, 3, 1, 3]\n\n", "entry_point": "minPath"} {"task": "code_ru_05957a846935d846a005e710", "id": "05957a846935d846a005e710", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/104", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef unique_digits(x):\n \"\"\"Given a list of positive integers x. return a sorted list of all \n elements that hasn't any even digit.\n\n Note: Returned list should be sorted in increasing order.\n \n For example:\n >>> unique_digits([15, 33, 1422, 1])\n [1, 15, 33]\n >>> unique_digits([152, 323, 1422, 10])\n []\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of filtering: the body keeps exactly those elements whose every decimal digit is odd (1,3,5,7,9) and excludes any element containing 0,2,4,6,8. Mental execution must give [1, 15, 33] for [15, 33, 1422, 1] and [] for [152, 323, 1422, 10]. Numbers like 10 or 20 that contain zero must be excluded. Duplicates in the input should be preserved, not deduplicated. A crash or non-termination on a valid list of positive integers scores 6 or below."}, {"id": "c2", "description": "Sorting and boundary cases: the result is returned in increasing order regardless of input order, and an empty input list returns []. Large values such as 999999999999 are handled correctly. String-based digit checks and arithmetic (%10) loops are equally acceptable. Extra guards that change outputs for valid inputs are penalized."}, {"id": "c3", "description": "Validity of the returned body: it is only the function body with consistent indentation that compiles when placed under the given signature. There is no repeated def line, no docstring, and no Markdown fences. If the first line's indentation breaks the body, note it and lower this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is roughly O(total digits + n log n), with clear logic such as all(int(d)%2 for d in str(n)) followed by sorted(). Readability is a minor factor compared with correctness. Commentary and unnecessary complexity earn no reward."}]}, "target_words": null, "reference": " odd_digit_elements = []\n for i in x:\n if all (int(c) % 2 == 1 for c in str(i)):\n odd_digit_elements.append(i)\n return sorted(odd_digit_elements)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/104", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef unique_digits(x):\n \"\"\"Given a list of positive integers x. return a sorted list of all \n elements that hasn't any even digit.\n\n Note: Returned list should be sorted in increasing order.\n \n For example:\n >>> unique_digits([15, 33, 1422, 1])\n [1, 15, 33]\n >>> unique_digits([152, 323, 1422, 10])\n []\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([15, 33, 1422, 1]) == [1, 15, 33]\n assert candidate([152, 323, 1422, 10]) == []\n assert candidate([12345, 2033, 111, 151]) == [111, 151]\n assert candidate([135, 103, 31]) == [31, 135]\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "unique_digits"} {"task": "code_ru_0caff09988c2c4ee4e4d2b25", "id": "0caff09988c2c4ee4e4d2b25", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/132", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef is_nested(string):\n '''\n Create a function that takes a string as input which contains only square brackets.\n The function should return True if and only if there is a valid subsequence of brackets \n where at least one bracket in the subsequence is nested.\n\n is_nested('[[]]') ➞ True\n is_nested('[]]]]]]][[[[[]') ➞ False\n is_nested('[][]') ➞ False\n is_nested('[]') ➞ False\n is_nested('[[][]]') ➞ True\n is_nested('[[]][[') ➞ True\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of nesting detection: the body returns True if and only if the string contains a subsequence of the form '[[]]' (two '[' followed later by two ']', which is the authoritative meaning of a nested valid subsequence). Mentally executing it on all six docstring examples gives True, False, False, False, True, True. Adversarial cases are also handled: '[[]' → False, '[]]' → False, ']][[' → False, '[][]' → False, '[[]]]' → True, and '[]][[]]' → True. A wrong value on a docstring example is severe. A wrong value on a rarer adversarial pattern is moderate."}, {"id": "c2", "description": "Boundary robustness: the empty string returns False without crashing. Strings with only '[' or only ']', single-character strings and very long strings all terminate without exceptions, such as an IndexError from popping an empty stack or indexing past the end. Any crash or non-termination on a valid bracket-only input caps this criterion at 6 or below."}, {"id": "c3", "description": "Valid returned body: the output is only a function body, with no signature, docstring or Markdown fences. Indentation is consistent, starting at function-body level, and the code is syntactically valid Python that parses when placed under the given signature. Broken first-line indentation or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach runs in O(n) or O(n log n), or is at least polynomial and reasonable. Examples include greedy scanning for two opens and then two closes, or index-list matching. It does not enumerate exponentially many subsequences. Variable names and structure are clear. Any correct alternative algorithm gets full credit, with no reward for extra commentary or unnecessary guards."}]}, "target_words": null, "reference": " opening_bracket_index = []\n closing_bracket_index = []\n for i in range(len(string)):\n if string[i] == '[':\n opening_bracket_index.append(i)\n else:\n closing_bracket_index.append(i)\n closing_bracket_index.reverse()\n cnt = 0\n i = 0\n l = len(closing_bracket_index)\n for idx in opening_bracket_index:\n if i < l and idx < closing_bracket_index[i]:\n cnt += 1\n i += 1\n return cnt >= 2\n\n \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/132", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef is_nested(string):\n '''\n Create a function that takes a string as input which contains only square brackets.\n The function should return True if and only if there is a valid subsequence of brackets \n where at least one bracket in the subsequence is nested.\n\n is_nested('[[]]') ➞ True\n is_nested('[]]]]]]][[[[[]') ➞ False\n is_nested('[][]') ➞ False\n is_nested('[]') ➞ False\n is_nested('[[][]]') ➞ True\n is_nested('[[]][[') ➞ True\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate('[[]]') == True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate('[]]]]]]][[[[[]') == False\n assert candidate('[][]') == False\n assert candidate(('[]')) == False\n assert candidate('[[[[]]]]') == True\n assert candidate('[]]]]]]]]]]') == False\n assert candidate('[][][[]]') == True\n assert candidate('[[]') == False\n assert candidate('[]]') == False\n assert candidate('[[]][[') == True\n assert candidate('[[][]]') == True\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate('') == False, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate('[[[[[[[[') == False\n assert candidate(']]]]]]]]') == False\n\n", "entry_point": "is_nested"} {"task": "code_ru_0d7604784124fdd3d5298e70", "id": "0d7604784124fdd3d5298e70", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/11", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nfrom typing import List\n\n\ndef string_xor(a: str, b: str) -> str:\n \"\"\" Input are two strings a and b consisting only of 1s and 0s.\n Perform binary XOR on these inputs and return result also as a string.\n >>> string_xor('010', '110')\n '100'\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: for equal-length inputs of 0s and 1s, the body returns a string where each character is '1' if the corresponding characters of a and b differ and '0' if they are the same. Check this by tracing the docstring example: '010','110' must give '100'."}, {"id": "c2", "description": "Boundary cases: empty strings return '' without crashing. Leading zeros are kept in the result, so string_xor('00','00') returns '00' and not '0' or ''. A solution that uses int(...,2) and bin() and drops leading zeros or the empty result counts as a moderate-to-severe correctness failure. A crash on an empty input scores 6 or below."}, {"id": "c3", "description": "Validity of the returned body: it contains only the function body, with no signature, no docstring and no Markdown fences. Its indentation is consistent, including the first line, and it parses as valid Python inside the def. Lower this criterion if indentation or syntax would break the function."}, {"id": "c4", "description": "Algorithmic suitability and readability: the body runs in linear time with a clear approach, such as a join over zip with a per-character comparison. It adds no extra guards or output changes beyond the specification, such as raising an error or padding unequal lengths in a way that alters the result on valid inputs."}]}, "target_words": null, "reference": " def xor(i, j):\n if i == j:\n return '0'\n else:\n return '1'\n\n return ''.join(xor(x, y) for x, y in zip(a, b))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/11", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "from typing import List\n\n\ndef string_xor(a: str, b: str) -> str:\n \"\"\" Input are two strings a and b consisting only of 1s and 0s.\n Perform binary XOR on these inputs and return result also as a string.\n >>> string_xor('010', '110')\n '100'\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('111000', '101010') == '010010'\n assert candidate('1', '1') == '0'\n assert candidate('0101', '0000') == '0101'\n", "entry_point": "string_xor"} {"task": "code_ru_1285642e9daaba6184198ac6", "id": "1285642e9daaba6184198ac6", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/63", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef fibfib(n: int):\n \"\"\"The FibFib number sequence is a sequence similar to the Fibbonacci sequnece that's defined as follows:\n fibfib(0) == 0\n fibfib(1) == 0\n fibfib(2) == 1\n fibfib(n) == fibfib(n-1) + fibfib(n-2) + fibfib(n-3).\n Please write a function to efficiently compute the n-th element of the fibfib number sequence.\n >>> fibfib(1)\n 0\n >>> fibfib(5)\n 4\n >>> fibfib(8)\n 24\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on base cases and docstring examples: returns 0 for n=0, 0 for n=1, 1 for n=2, and 4 for n=5, 24 for n=8 (also e.g. fibfib(3)=1, fibfib(4)=2), with no off-by-one error in the recurrence fibfib(n-1)+fibfib(n-2)+fibfib(n-3)."}, {"id": "c2", "description": "Efficiency: computes the n-th term in linear time (iterative rolling variables, memoization, or matrix exponentiation). It must not use naive exponential recursion. It should handle large n such as 1000+ without hanging and without hitting the recursion limit. A crash or effective non-termination on a large in-domain n scores 6 or below."}, {"id": "c3", "description": "Boundary handling: small n (0, 1, 2) are returned directly or fall out of the loop correctly without index errors. No extra guards change outputs for valid non-negative n. Negative n is outside the specified domain and should not be penalized unless it causes non-termination on in-domain inputs."}, {"id": "c4", "description": "Validity and readability: the output is only the function body, with consistent indentation that would parse when placed under the signature. There is no repeated def or docstring and no Markdown fences. Variable names and structure are clear. A broken first-line indentation or syntax error lowers this criterion substantially."}]}, "target_words": null, "reference": " if n == 0:\n return 0\n if n == 1:\n return 0\n if n == 2:\n return 1\n return fibfib(n - 1) + fibfib(n - 2) + fibfib(n - 3)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/63", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef fibfib(n: int):\n \"\"\"The FibFib number sequence is a sequence similar to the Fibbonacci sequnece that's defined as follows:\n fibfib(0) == 0\n fibfib(1) == 0\n fibfib(2) == 1\n fibfib(n) == fibfib(n-1) + fibfib(n-2) + fibfib(n-3).\n Please write a function to efficiently compute the n-th element of the fibfib number sequence.\n >>> fibfib(1)\n 0\n >>> fibfib(5)\n 4\n >>> fibfib(8)\n 24\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(2) == 1\n assert candidate(1) == 0\n assert candidate(5) == 4\n assert candidate(8) == 24\n assert candidate(10) == 81\n assert candidate(12) == 274\n assert candidate(14) == 927\n\n", "entry_point": "fibfib"} {"task": "code_ru_15fe40001d984fa7163c3548", "id": "15fe40001d984fa7163c3548", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/111", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef histogram(test):\n \"\"\"Given a string representing a space separated lowercase letters, return a dictionary\n of the letter with the most repetition and containing the corresponding count.\n If several letters have the same occurrence, return all of them.\n \n Example:\n histogram('a b c') == {'a': 1, 'b': 1, 'c': 1}\n histogram('a b b a') == {'a': 2, 'b': 2}\n histogram('a b c a b') == {'a': 2, 'b': 2}\n histogram('b b b b a') == {'b': 4}\n histogram('') == {}\n\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields exactly {'a': 1, 'b': 1, 'c': 1}, {'a': 2, 'b': 2}, {'a': 2, 'b': 2}, {'b': 4} for the given inputs, including all tied letters with the maximum count and excluding letters with lower counts."}, {"id": "c2", "description": "Boundary handling: empty string '' returns {} without crashing (e.g., no max() on an empty sequence, no counting '' as a letter). Inputs with extra or consecutive spaces should not produce an empty-string key, or this should be treated only as a minor issue since the specification implies single spaces. A crash on '' is severe (score 6 or below)."}, {"id": "c3", "description": "Validity of returned body: the code is only a function body with consistent, correct indentation and valid Python syntax, without a repeated signature, docstring, or Markdown fences. Any broken first-line indentation or syntax error is noted and lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: the code counts letters in roughly linear time (e.g., dict or Counter over split tokens rather than repeated list.count in a nested loop, though that is acceptable for correctness). It uses clear variable names and adds no extra guards that change outputs."}]}, "target_words": null, "reference": " dict1={}\n list1=test.split(\" \")\n t=0\n\n for i in list1:\n if(list1.count(i)>t) and i!='':\n t=list1.count(i)\n if t>0:\n for i in list1:\n if(list1.count(i)==t):\n \n dict1[i]=t\n return dict1\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/111", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef histogram(test):\n \"\"\"Given a string representing a space separated lowercase letters, return a dictionary\n of the letter with the most repetition and containing the corresponding count.\n If several letters have the same occurrence, return all of them.\n \n Example:\n histogram('a b c') == {'a': 1, 'b': 1, 'c': 1}\n histogram('a b b a') == {'a': 2, 'b': 2}\n histogram('a b c a b') == {'a': 2, 'b': 2}\n histogram('b b b b a') == {'b': 4}\n histogram('') == {}\n\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate('a b b a') == {'a':2,'b': 2}, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate('a b c a b') == {'a': 2, 'b': 2}, \"This prints if this assert fails 2 (good for debugging!)\"\n assert candidate('a b c d g') == {'a': 1, 'b': 1, 'c': 1, 'd': 1, 'g': 1}, \"This prints if this assert fails 3 (good for debugging!)\"\n assert candidate('r t g') == {'r': 1,'t': 1,'g': 1}, \"This prints if this assert fails 4 (good for debugging!)\"\n assert candidate('b b b b a') == {'b': 4}, \"This prints if this assert fails 5 (good for debugging!)\"\n assert candidate('r t g') == {'r': 1,'t': 1,'g': 1}, \"This prints if this assert fails 6 (good for debugging!)\"\n \n \n # Check some edge cases that are easy to work out by hand.\n assert candidate('') == {}, \"This prints if this assert fails 7 (also good for debugging!)\"\n assert candidate('a') == {'a': 1}, \"This prints if this assert fails 8 (also good for debugging!)\"\n\n", "entry_point": "histogram"} {"task": "code_ru_17863015a4533999080c89ea", "id": "17863015a4533999080c89ea", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/56", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef correct_bracketing(brackets: str):\n \"\"\" brackets is a string of \"<\" and \">\".\n return True if every opening bracket has a corresponding closing bracket.\n\n >>> correct_bracketing(\"<\")\n False\n >>> correct_bracketing(\"<>\")\n True\n >>> correct_bracketing(\"<<><>>\")\n True\n >>> correct_bracketing(\"><<>\")\n False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of balance logic: the body tracks nesting depth (counter or stack), returns False immediately if depth ever goes negative (e.g. \"><<>\" → False, \">\" → False), and returns True only if final depth is zero (\"<\" → False, \"<>\" → True, \"<<><>>\" → True). Mentally executing all docstring examples yields the stated results; a solution that only compares counts of '<' and '>' without prefix checking fails \"><<>\"-style inputs and should score low."}, {"id": "c2", "description": "Boundary cases: empty string returns True; strings of only '<' or only '>' return False; long inputs (e.g. 10^5 characters) are handled without recursion depth errors or quadratic behavior such as repeated str.replace('<>','') loops. Any crash or non-termination on a valid bracket string scores 6 or below."}, {"id": "c3", "description": "Output validity: the response is only the function body (no signature, docstring, or Markdown fences), consistently indented so it parses when placed under the given def; first-line indentation errors or syntax errors lower this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: uses a single O(n) pass with O(1) or O(n) extra space, clear variable names and straightforward control flow; does not add guards or behaviors that change outputs for valid inputs (e.g. raising on other characters is not required and must not alter results on '<'/'>' strings)."}]}, "target_words": null, "reference": " depth = 0\n for b in brackets:\n if b == \"<\":\n depth += 1\n else:\n depth -= 1\n if depth < 0:\n return False\n return depth == 0\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/56", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef correct_bracketing(brackets: str):\n \"\"\" brackets is a string of \"<\" and \">\".\n return True if every opening bracket has a corresponding closing bracket.\n\n >>> correct_bracketing(\"<\")\n False\n >>> correct_bracketing(\"<>\")\n True\n >>> correct_bracketing(\"<<><>>\")\n True\n >>> correct_bracketing(\"><<>\")\n False\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(\"<>\")\n assert candidate(\"<<><>>\")\n assert candidate(\"<><><<><>><>\")\n assert candidate(\"<><><<<><><>><>><<><><<>>>\")\n assert not candidate(\"<<<><>>>>\")\n assert not candidate(\"><<>\")\n assert not candidate(\"<\")\n assert not candidate(\"<<<<\")\n assert not candidate(\">\")\n assert not candidate(\"<<>\")\n assert not candidate(\"<><><<><>><>><<>\")\n assert not candidate(\"<><><<><>><>>><>\")\n\n", "entry_point": "correct_bracketing"} {"task": "code_ru_26fac9d2a55d0d8c47ef10ac", "id": "26fac9d2a55d0d8c47ef10ac", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/25", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nfrom typing import List\n\n\ndef factorize(n: int) -> List[int]:\n \"\"\" Return list of prime factors of given integer in the order from smallest to largest.\n Each of the factors should be listed number of times corresponding to how many times it appeares in factorization.\n Input number should be equal to the product of all factors\n >>> factorize(8)\n [2, 2, 2]\n >>> factorize(25)\n [5, 5]\n >>> factorize(70)\n [2, 5, 7]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: the body returns [2, 2, 2] for 8, [5, 5] for 25 and [2, 5, 7] for 70. Factors are in nondecreasing order, repeated by multiplicity, and their product equals n."}, {"id": "c2", "description": "Boundary and adversarial inputs. n=1 returns [] and must not loop forever. A prime n (e.g. 2, 13, 97) returns [n]. A large prime or semiprime, such as 999983 or 2*999983, must keep the leftover factor greater than sqrt(n) and must not drop it. Prime powers and mixed multiplicities (e.g. 360 -> [2,2,2,3,3,5]) must be correct. Non-termination or a crash for any n>=1 scores 6 or below. Behavior for n<=0 is not specified and should not be penalized unless the code hangs."}, {"id": "c3", "description": "Algorithmic suitability: uses trial division up to sqrt of the remaining n, or an equally efficient correct method, so that inputs around 1e12 finish promptly. Iterating divisors up to n itself on a large prime counts as a performance weakness. Any correct algorithm gets full credit."}, {"id": "c4", "description": "Validity and readability. The output is only a function body, with no signature, docstring or Markdown fences. It has consistent indentation that is valid when placed under the signature, with no first-line indentation errors. It is syntactically correct Python, and its variable names and loop structure are clear. Extra commentary and guards that change outputs earn no credit."}]}, "target_words": null, "reference": " import math\n fact = []\n i = 2\n while i <= int(math.sqrt(n) + 1):\n if n % i == 0:\n fact.append(i)\n n //= i\n else:\n i += 1\n\n if n > 1:\n fact.append(n)\n return fact\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/25", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "from typing import List\n\n\ndef factorize(n: int) -> List[int]:\n \"\"\" Return list of prime factors of given integer in the order from smallest to largest.\n Each of the factors should be listed number of times corresponding to how many times it appeares in factorization.\n Input number should be equal to the product of all factors\n >>> factorize(8)\n [2, 2, 2]\n >>> factorize(25)\n [5, 5]\n >>> factorize(70)\n [2, 5, 7]\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate(2) == [2]\n assert candidate(4) == [2, 2]\n assert candidate(8) == [2, 2, 2]\n assert candidate(3 * 19) == [3, 19]\n assert candidate(3 * 19 * 3 * 19) == [3, 3, 19, 19]\n assert candidate(3 * 19 * 3 * 19 * 3 * 19) == [3, 3, 3, 19, 19, 19]\n assert candidate(3 * 19 * 19 * 19) == [3, 19, 19, 19]\n assert candidate(3 * 2 * 3) == [2, 3, 3]\n", "entry_point": "factorize"} {"task": "code_ru_3d0f31e2bfa770efcfa77ab3", "id": "3d0f31e2bfa770efcfa77ab3", "domain": "code", "language": "ru", "split": "validation", "group": "ru:HumanEval/86", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef anti_shuffle(s):\n \"\"\"\n Write a function that takes a string and returns an ordered version of it.\n Ordered version of string, is a string where all words (separated by space)\n are replaced by a new word where all the characters arranged in\n ascending order based on ascii value.\n Note: You should keep the order of words and blank spaces in the sentence.\n\n For example:\n anti_shuffle('Hi') returns 'Hi'\n anti_shuffle('hello') returns 'ehllo'\n anti_shuffle('Hello World!!!') returns 'Hello !!!Wdlor'\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields 'Hi' for 'Hi', 'ehllo' for 'hello', and 'Hello !!!Wdlor' for 'Hello World!!!'. Each word's characters must be sorted by ASCII value, so uppercase letters come before lowercase letters and '!' comes before letters. Any wrong output here is a major correctness failure."}, {"id": "c2", "description": "Spacing preservation: the body keeps the exact positions of spaces, including multiple consecutive spaces, leading spaces and trailing spaces. For example, 'b a c ' stays 'b a c '. The empty string returns ''. A correct solution typically splits on ' ' (not split() with no argument) and joins with ' ', or processes characters in an equivalent way. Collapsing or stripping spaces is a wrong value on a plausible input and counts as a moderate-to-significant deduction. A crash on '' is severe: score 6 or below."}, {"id": "c3", "description": "Validity of the returned body: the output is only the function body, without the signature, the docstring or Markdown fences. Indentation is consistent and valid when placed under the def. It is syntactically correct Python and returns a str. Lower this criterion if the first line's indentation breaks the body, or if extraneous text is present."}, {"id": "c4", "description": "Algorithmic suitability and readability: the body uses a straightforward approach, such as sorted() per word with ''.join(). It runs in roughly O(n log n) time. It is clear and concise and adds no guards that change outputs. Any correct alternative algorithm deserves full credit, and added commentary or complexity is not rewarded."}]}, "target_words": null, "reference": " return ' '.join([''.join(sorted(list(i))) for i in s.split(' ')])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/86", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef anti_shuffle(s):\n \"\"\"\n Write a function that takes a string and returns an ordered version of it.\n Ordered version of string, is a string where all words (separated by space)\n are replaced by a new word where all the characters arranged in\n ascending order based on ascii value.\n Note: You should keep the order of words and blank spaces in the sentence.\n\n For example:\n anti_shuffle('Hi') returns 'Hi'\n anti_shuffle('hello') returns 'ehllo'\n anti_shuffle('Hello World!!!') returns 'Hello !!!Wdlor'\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate('Hi') == 'Hi'\n assert candidate('hello') == 'ehllo'\n assert candidate('number') == 'bemnru'\n assert candidate('abcd') == 'abcd'\n assert candidate('Hello World!!!') == 'Hello !!!Wdlor'\n assert candidate('') == ''\n assert candidate('Hi. My name is Mister Robot. How are you?') == '.Hi My aemn is Meirst .Rboot How aer ?ouy'\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "anti_shuffle"} {"task": "code_ru_3f6051877583b9097d592069", "id": "3f6051877583b9097d592069", "domain": "code", "language": "ru", "split": "validation", "group": "ru:HumanEval/36", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef fizz_buzz(n: int):\n \"\"\"Return the number of times the digit 7 appears in integers less than n which are divisible by 11 or 13.\n >>> fizz_buzz(50)\n 0\n >>> fizz_buzz(78)\n 2\n >>> fizz_buzz(79)\n 3\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core semantics: the body counts occurrences of the digit '7' (every occurrence, so 77 contributes 2) across all integers i with 0 <= i < n that are divisible by 11 or 13. Numbers divisible by both, such as 143, are counted once, not twice. Mentally executing it yields fizz_buzz(50)=0, fizz_buzz(78)=2 (from 77) and fizz_buzz(79)=3 (adding 78)."}, {"id": "c2", "description": "Boundary handling: the range excludes n itself. n=0, n=1 and negative n return 0 without crashing or looping. Including 0 in the range is harmless because 0 contains no 7. A crash or non-termination on any integer n caps this criterion at 6 or below."}, {"id": "c3", "description": "Validity of the returned format: the output is only the function body, with no signature, no docstring and no Markdown fences. Indentation is consistent and the body parses as valid Python inside the given def, including the first line. It returns an int and does not print."}, {"id": "c4", "description": "Algorithmic suitability and readability: it uses a straightforward O(n·digits) approach, such as iterating over the range and summing str(i).count('7') or an equivalent digit count, or a correct alternative. The code is concise and clear and adds no guards or special cases that change outputs."}]}, "target_words": null, "reference": " ns = []\n for i in range(n):\n if i % 11 == 0 or i % 13 == 0:\n ns.append(i)\n s = ''.join(list(map(str, ns)))\n ans = 0\n for c in s:\n ans += (c == '7')\n return ans\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/36", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef fizz_buzz(n: int):\n \"\"\"Return the number of times the digit 7 appears in integers less than n which are divisible by 11 or 13.\n >>> fizz_buzz(50)\n 0\n >>> fizz_buzz(78)\n 2\n >>> fizz_buzz(79)\n 3\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(50) == 0\n assert candidate(78) == 2\n assert candidate(79) == 3\n assert candidate(100) == 3\n assert candidate(200) == 6\n assert candidate(4000) == 192\n assert candidate(10000) == 639\n assert candidate(100000) == 8026\n\n", "entry_point": "fizz_buzz"} {"task": "code_ru_4456264cb70a760738d05aff", "id": "4456264cb70a760738d05aff", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/78", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef hex_key(num):\n \"\"\"You have been tasked to write a function that receives \n a hexadecimal number as a string and counts the number of hexadecimal \n digits that are primes (prime number, or a prime, is a natural number \n greater than 1 that is not a product of two smaller natural numbers).\n Hexadecimal digits are 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, F.\n Prime numbers are 2, 3, 5, 7, 11, 13, 17,...\n So you have to determine a number of the following digits: 2, 3, 5, 7, \n B (=decimal 11), D (=decimal 13).\n Note: you may assume the input is always correct or empty string, \n and symbols A,B,C,D,E,F are always uppercase.\n Examples:\n For num = \"AB\" the output should be 1.\n For num = \"1077E\" the output should be 2.\n For num = \"ABED1A33\" the output should be 4.\n For num = \"123456789ABCDEF0\" the output should be 6.\n For num = \"2020\" the output should be 2.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body counts every character of num that belongs to the set {'2','3','5','7','B','D'}, counting duplicates. It must return 1 for \"AB\", 2 for \"1077E\", 4 for \"ABED1A33\", 6 for \"123456789ABCDEF0\" and 2 for \"2020\". Digits 0, 1, 4, 6, 8, 9, A, C, E and F must not be counted. Any digit misclassified as prime or non-prime is a major failure."}, {"id": "c2", "description": "Boundary cases: an empty string returns 0 without crashing. A string containing only non-prime digits returns 0. A long string is handled correctly. A crash or non-termination on any valid input, including the empty string, caps this criterion at 6 or below. The body must not add extra handling, such as lowercase conversion or validation that raises, that changes outputs on valid inputs."}, {"id": "c3", "description": "Validity and format: the returned text is only a function body. It does not repeat the def line or the docstring and contains no Markdown fences. Its indentation is consistent, with the first line included, so it forms syntactically valid Python once placed under the signature. The body returns an int."}, {"id": "c4", "description": "Algorithm and readability: the body uses a simple linear pass, such as set membership or a sum over characters, with clear names. Readability is a minor factor compared with correctness. A correct alternative approach gets full credit."}]}, "target_words": null, "reference": " primes = ('2', '3', '5', '7', 'B', 'D')\n total = 0\n for i in range(0, len(num)):\n if num[i] in primes:\n total += 1\n return total\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/78", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef hex_key(num):\n \"\"\"You have been tasked to write a function that receives \n a hexadecimal number as a string and counts the number of hexadecimal \n digits that are primes (prime number, or a prime, is a natural number \n greater than 1 that is not a product of two smaller natural numbers).\n Hexadecimal digits are 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A, B, C, D, E, F.\n Prime numbers are 2, 3, 5, 7, 11, 13, 17,...\n So you have to determine a number of the following digits: 2, 3, 5, 7, \n B (=decimal 11), D (=decimal 13).\n Note: you may assume the input is always correct or empty string, \n and symbols A,B,C,D,E,F are always uppercase.\n Examples:\n For num = \"AB\" the output should be 1.\n For num = \"1077E\" the output should be 2.\n For num = \"ABED1A33\" the output should be 4.\n For num = \"123456789ABCDEF0\" the output should be 6.\n For num = \"2020\" the output should be 2.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"AB\") == 1, \"First test error: \" + str(candidate(\"AB\")) \n assert candidate(\"1077E\") == 2, \"Second test error: \" + str(candidate(\"1077E\")) \n assert candidate(\"ABED1A33\") == 4, \"Third test error: \" + str(candidate(\"ABED1A33\")) \n assert candidate(\"2020\") == 2, \"Fourth test error: \" + str(candidate(\"2020\")) \n assert candidate(\"123456789ABCDEF0\") == 6, \"Fifth test error: \" + str(candidate(\"123456789ABCDEF0\")) \n assert candidate(\"112233445566778899AABBCCDDEEFF00\") == 12, \"Sixth test error: \" + str(candidate(\"112233445566778899AABBCCDDEEFF00\")) \n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([]) == 0\n\n", "entry_point": "hex_key"} {"task": "code_ru_44f3277d44bcda7c864042bf", "id": "44f3277d44bcda7c864042bf", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/47", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef median(l: list):\n \"\"\"Return median of elements in the list l.\n >>> median([3, 1, 2, 4, 5])\n 3\n >>> median([-10, 4, 6, 1000, 10, 20])\n 15.0\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Computes the standard median correctly. It sorts a copy of the list or uses an equivalent selection method. For odd length it returns the middle element, e.g. median([3, 1, 2, 4, 5]) == 3. For even length it returns the arithmetic mean of the two middle elements as a float, e.g. 8.0 for [-10, 4, 6, 1000, 10, 20]. The docstring's 15.0 is a known erratum and should not be reproduced through ad-hoc logic. Hardcoding or special-casing the examples is a severe failure."}, {"id": "c2", "description": "Handles boundary inputs within the domain of non-empty numeric lists. A single element returns that element. Two elements return their mean. Negative values, duplicates, floats and already-sorted or reverse-sorted input all work. The function does not mutate the caller's list, or any in-place sort is harmless to the result. Empty-list behavior is unspecified, so a raised exception is acceptable. However, an infinite loop or a crash on any non-empty list caps this criterion at 6 or below."}, {"id": "c3", "description": "The returned body is syntactically valid Python when placed under the given signature. It has consistent indentation starting at function-body level and includes no repeated def line, docstring or Markdown fences. A body whose first-line indentation breaks parsing scores low."}, {"id": "c4", "description": "Uses a suitable algorithm, such as O(n log n) sorting or an O(n) selection method. The code is readable, with clear index arithmetic like n // 2 and n // 2 - 1. It contains no needless complexity and no extra guards that change outputs for valid inputs."}]}, "target_words": null, "reference": " l = sorted(l)\n if len(l) % 2 == 1:\n return l[len(l) // 2]\n else:\n return (l[len(l) // 2 - 1] + l[len(l) // 2]) / 2.0\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/47", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef median(l: list):\n \"\"\"Return median of elements in the list l.\n >>> median([3, 1, 2, 4, 5])\n 3\n >>> median([-10, 4, 6, 1000, 10, 20])\n 15.0\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([3, 1, 2, 4, 5]) == 3\n assert candidate([-10, 4, 6, 1000, 10, 20]) == 8.0\n assert candidate([5]) == 5\n assert candidate([6, 5]) == 5.5\n assert candidate([8, 1, 3, 9, 9, 2, 7]) == 7 \n\n", "entry_point": "median"} {"task": "code_ru_4ea6795070c7a869efe623ba", "id": "4ea6795070c7a869efe623ba", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/34", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef unique(l: list):\n \"\"\"Return sorted unique elements in a list\n >>> unique([5, 3, 5, 2, 3, 3, 9, 0, 123])\n [0, 2, 3, 5, 9, 123]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns a new list of the distinct elements of l in ascending order, e.g. sorted(set(l)). Mentally executing it on [5, 3, 5, 2, 3, 3, 9, 0, 123] must yield exactly [0, 2, 3, 5, 9, 123]. The return value must be a list, not a set, a generator or None. Returning an unsorted or still-duplicated result is a major failure."}, {"id": "c2", "description": "Boundary cases: the empty list returns []. A single-element list returns that element in a list. A list of all-identical values returns one element. Negative numbers sort correctly, e.g. [-1, -5, -1] gives [-5, -1]. Mixed int and float values that compare equal (1 and 1.0) do not crash. Raising or hanging on any of these inputs is severe and scores 6 or below. Mutating the input list in place is a moderate flaw."}, {"id": "c3", "description": "Validity and format: the output is only a function body, with no def line, no docstring and no Markdown fences. It has consistent indentation, including the first line, so it parses when placed under the given signature. A syntax or indentation break lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is O(n log n) or better, such as set plus sort. Quadratic membership checks are acceptable but less suitable. The code is concise and clear. Do not reward extra commentary, or guards that change outputs."}]}, "target_words": null, "reference": " return sorted(list(set(l)))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/34", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef unique(l: list):\n \"\"\"Return sorted unique elements in a list\n >>> unique([5, 3, 5, 2, 3, 3, 9, 0, 123])\n [0, 2, 3, 5, 9, 123]\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([5, 3, 5, 2, 3, 3, 9, 0, 123]) == [0, 2, 3, 5, 9, 123]\n\n", "entry_point": "unique"} {"task": "code_ru_572ff36e0755cce3c91731ea", "id": "572ff36e0755cce3c91731ea", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/150", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef x_or_y(n, x, y):\n \"\"\"A simple program which should return the value of x if n is \n a prime number and should return the value of y otherwise.\n\n Examples:\n for x_or_y(7, 34, 12) == 34\n for x_or_y(15, 8, 5) == 5\n \n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Primality is decided correctly: returns x for primes (e.g. 2, 3, 7, 13) and y for composites (e.g. 4, 9, 15, 25, perfect squares of primes). This includes the docstring examples x_or_y(7,34,12)==34 and x_or_y(15,8,5)==5."}, {"id": "c2", "description": "Boundary inputs are handled: n=1, n=0 and negative n all return y because they are not prime, and n=2 returns x. The code does not crash or loop forever on any of these inputs."}, {"id": "c3", "description": "The algorithm is suitable. Trial division up to sqrt(n), or an equivalent correct method, finishes promptly for large n such as 10**9+7 rather than depending on an O(n) loop or worse. A correct alternative method gets full credit."}, {"id": "c4", "description": "The output is a valid function body only, with no signature, docstring or Markdown. Indentation is consistent and fits inside the def, including the first line, and the code is readable with no extra guards that change outputs."}]}, "target_words": null, "reference": " if n == 1:\n return y\n for i in range(2, n):\n if n % i == 0:\n return y\n break\n else:\n return x\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/150", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef x_or_y(n, x, y):\n \"\"\"A simple program which should return the value of x if n is \n a prime number and should return the value of y otherwise.\n\n Examples:\n for x_or_y(7, 34, 12) == 34\n for x_or_y(15, 8, 5) == 5\n \n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(7, 34, 12) == 34\n assert candidate(15, 8, 5) == 5\n assert candidate(3, 33, 5212) == 33\n assert candidate(1259, 3, 52) == 3\n assert candidate(7919, -1, 12) == -1\n assert candidate(3609, 1245, 583) == 583\n assert candidate(91, 56, 129) == 129\n assert candidate(6, 34, 1234) == 1234\n \n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(1, 2, 0) == 0\n assert candidate(2, 2, 0) == 2\n\n", "entry_point": "x_or_y"} {"task": "code_ru_59f67b2d99d5b07ec6635555", "id": "59f67b2d99d5b07ec6635555", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/123", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef get_odd_collatz(n):\n \"\"\"\n Given a positive integer n, return a sorted list that has the odd numbers in collatz sequence.\n\n The Collatz conjecture is a conjecture in mathematics that concerns a sequence defined\n as follows: start with any positive integer n. Then each term is obtained from the \n previous term as follows: if the previous term is even, the next term is one half of \n the previous term. If the previous term is odd, the next term is 3 times the previous\n term plus 1. The conjecture is that no matter what value of n, the sequence will always reach 1.\n\n Note: \n 1. Collatz(1) is [1].\n 2. returned list sorted in increasing order.\n\n For example:\n get_odd_collatz(5) returns [1, 5] # The collatz sequence for 5 is [5, 16, 8, 4, 2, 1], so the odd numbers are only 1, and 5.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core examples: mentally executing the body gives get_odd_collatz(5) == [1, 5] and get_odd_collatz(1) == [1]. It also handles values like 3 ([1,3,5,...]) and 6 ([1,3,5]) correctly: it follows the sequence until it reaches 1, collects each odd term once, and returns the list sorted ascending. Wrong outputs on these inputs are a major failure."}, {"id": "c2", "description": "Termination and boundary behavior: the loop terminates for all positive integers, including n=1, even starting values such as powers of two (e.g. 16 -> [1]), and large n such as 27 or 10**6. It does not loop forever (e.g. by checking a wrong stop condition after 1), crash, or omit the terminal 1. Non-termination or a crash on any positive n caps this criterion at 6 or below. It does not penalize missing handling of n<=0, which is outside the stated domain."}, {"id": "c3", "description": "Output fidelity: the result contains no duplicate odd values and no even values. It correctly includes the starting n when n is odd. It uses integer arithmetic (// or equivalent) rather than float division that could produce floats or lose precision for large n. The returned value is a list of ints in increasing order."}, {"id": "c4", "description": "Validity and readability: the returned text is only a function body with consistent indentation, starting at the proper body indentation level. It is syntactically valid Python with no repeated signature, no docstring, and no Markdown fences. The code is clear and appropriately simple, with a direct simulation in O(sequence length) plus a sort. Broken first-line indentation or syntax errors lower this criterion substantially. Cosmetic style weighs less than correctness."}]}, "target_words": null, "reference": " if n%2==0:\n odd_collatz = [] \n else:\n odd_collatz = [n]\n while n > 1:\n if n % 2 == 0:\n n = n/2\n else:\n n = n*3 + 1\n \n if n%2 == 1:\n odd_collatz.append(int(n))\n\n return sorted(odd_collatz)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/123", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef get_odd_collatz(n):\n \"\"\"\n Given a positive integer n, return a sorted list that has the odd numbers in collatz sequence.\n\n The Collatz conjecture is a conjecture in mathematics that concerns a sequence defined\n as follows: start with any positive integer n. Then each term is obtained from the \n previous term as follows: if the previous term is even, the next term is one half of \n the previous term. If the previous term is odd, the next term is 3 times the previous\n term plus 1. The conjecture is that no matter what value of n, the sequence will always reach 1.\n\n Note: \n 1. Collatz(1) is [1].\n 2. returned list sorted in increasing order.\n\n For example:\n get_odd_collatz(5) returns [1, 5] # The collatz sequence for 5 is [5, 16, 8, 4, 2, 1], so the odd numbers are only 1, and 5.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(14) == [1, 5, 7, 11, 13, 17]\n assert candidate(5) == [1, 5]\n assert candidate(12) == [1, 3, 5], \"This prints if this assert fails 1 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(1) == [1], \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "get_odd_collatz"} {"task": "code_ru_5af73518e8caf1516c903d26", "id": "5af73518e8caf1516c903d26", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/71", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef triangle_area(a, b, c):\n '''\n Given the lengths of the three sides of a triangle. Return the area of\n the triangle rounded to 2 decimal points if the three sides form a valid triangle. \n Otherwise return -1\n Three sides make a valid triangle when the sum of any two sides is greater \n than the third side.\n Example:\n triangle_area(3, 4, 5) == 6.00\n triangle_area(1, 2, 10) == -1\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Validity check is correct: returns -1 exactly when any pair of sides fails the strict inequality (sum of two sides greater than the third), so degenerate triangles like (1, 2, 3) return -1, and (1, 2, 10) returns -1. Zero or negative sides are rejected as a consequence of the strict triangle inequality, with no crash."}, {"id": "c2", "description": "For valid triangles, the area is computed correctly, for example with Heron's formula using semi-perimeter s=(a+b+c)/2 and sqrt(s(s-a)(s-b)(s-c)). The result is rounded to 2 decimals with round(..., 2), so (3, 4, 5) returns 6.0 and (1, 1, 1) returns 0.43. No integer division or truncation errors occur."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation that fits directly under the given signature. It contains no repeated def line, docstring, or Markdown fences. Any import (e.g. math) is placed inside the body or avoided by using ** 0.5."}, {"id": "c4", "description": "The code is concise and readable, with constant-time logic. It has no extra guards or side effects that change outputs for inputs in the spec, such as raising exceptions, printing, or returning values other than -1 or the rounded float."}]}, "target_words": null, "reference": " if a + b <= c or a + c <= b or b + c <= a:\n return -1 \n s = (a + b + c)/2 \n area = (s * (s - a) * (s - b) * (s - c)) ** 0.5\n area = round(area, 2)\n return area\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/71", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef triangle_area(a, b, c):\n '''\n Given the lengths of the three sides of a triangle. Return the area of\n the triangle rounded to 2 decimal points if the three sides form a valid triangle. \n Otherwise return -1\n Three sides make a valid triangle when the sum of any two sides is greater \n than the third side.\n Example:\n triangle_area(3, 4, 5) == 6.00\n triangle_area(1, 2, 10) == -1\n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(3, 4, 5) == 6.00, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(1, 2, 10) == -1\n assert candidate(4, 8, 5) == 8.18\n assert candidate(2, 2, 2) == 1.73\n assert candidate(1, 2, 3) == -1\n assert candidate(10, 5, 7) == 16.25\n assert candidate(2, 6, 3) == -1\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(1, 1, 1) == 0.43, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate(2, 2, 10) == -1\n\n", "entry_point": "triangle_area"} {"task": "code_ru_5c4eeb49abe3759f562020c9", "id": "5c4eeb49abe3759f562020c9", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/6", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nfrom typing import List\n\n\ndef parse_nested_parens(paren_string: str) -> List[int]:\n \"\"\" Input to this function is a string represented multiple groups for nested parentheses separated by spaces.\n For each of the group, output the deepest level of nesting of parentheses.\n E.g. (()()) has maximum two levels of nesting while ((())) has three.\n\n >>> parse_nested_parens('(()()) ((())) () ((())()())')\n [2, 3, 1, 3]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on the docstring example: for '(()()) ((())) () ((())()())' the body returns exactly [2, 3, 1, 3]. It computes each group's maximum depth by tracking the current depth (increment on '(', decrement on ')') and recording the peak, or by any equivalent correct method."}, {"id": "c2", "description": "Boundary handling: the body splits groups on spaces without producing entries for empty tokens. Multiple, leading or trailing spaces do not add spurious 0 values, for example by filtering empty strings or using split(). An empty input string returns [] rather than [0] or a crash. Depth is reset for each group. Loss of [] on empty input is a moderate error; a crash on a well-formed input is severe and scores 6 or below."}, {"id": "c3", "description": "Validity as a returned body: the code is only the function body, without repeating the signature, docstring or Markdown fences. It is consistently indented so that it is syntactically valid when placed under the given def, including the first line's indentation. It uses only available names (List is already imported) and terminates on all inputs."}, {"id": "c4", "description": "Algorithmic suitability and readability: it runs in linear time over the string, with no exponential or repeated-rescan approaches. Variable names and structure are clear and concise. There are no extra guards or validation that change outputs for valid inputs. Excess commentary is not rewarded."}]}, "target_words": null, "reference": " def parse_paren_group(s):\n depth = 0\n max_depth = 0\n for c in s:\n if c == '(':\n depth += 1\n max_depth = max(depth, max_depth)\n else:\n depth -= 1\n\n return max_depth\n\n return [parse_paren_group(x) for x in paren_string.split(' ') if x]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/6", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "from typing import List\n\n\ndef parse_nested_parens(paren_string: str) -> List[int]:\n \"\"\" Input to this function is a string represented multiple groups for nested parentheses separated by spaces.\n For each of the group, output the deepest level of nesting of parentheses.\n E.g. (()()) has maximum two levels of nesting while ((())) has three.\n\n >>> parse_nested_parens('(()()) ((())) () ((())()())')\n [2, 3, 1, 3]\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('(()()) ((())) () ((())()())') == [2, 3, 1, 3]\n assert candidate('() (()) ((())) (((())))') == [1, 2, 3, 4]\n assert candidate('(()(())((())))') == [4]\n", "entry_point": "parse_nested_parens"} {"task": "code_ru_64f4e222024b92df0e795379", "id": "64f4e222024b92df0e795379", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/38", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef encode_cyclic(s: str):\n \"\"\"\n returns encoded string by cycling groups of three characters.\n \"\"\"\n # split string to groups. Each of length 3.\n groups = [s[(3 * i):min((3 * i + 3), len(s))] for i in range((len(s) + 2) // 3)]\n # cycle elements in each group. Unless group has fewer elements than 3.\n groups = [(group[1:] + group[0]) if len(group) == 3 else group for group in groups]\n return \"\".join(groups)\n\n\ndef decode_cyclic(s: str):\n \"\"\"\n takes as input string encoded with encode_cyclic function. Returns decoded string.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of inversion: for every string s, decode_cyclic(encode_cyclic(s)) == s. Each full 3-character group is rotated right by one (group[-1] + group[:-1]), or encode is applied twice. Trailing groups of length 1 or 2 are left unchanged. Check this mentally on 'abc'->'bca'->'abc', 'abcdefg', and 'abcde'. A wrong rotation direction or rotating short groups is a major failure."}, {"id": "c2", "description": "Boundary cases: the empty string returns '' without error, and strings of length 1, 2, and exact multiples of 3 decode correctly. Any crash (e.g., IndexError on short groups) or non-termination on valid string input scores 6 or below."}, {"id": "c3", "description": "Validity of the returned body: it is only the function body, with no signature, docstring, or Markdown fences. It has consistent indentation that parses as the body of decode_cyclic, and it ends with a return statement. A broken first-line indentation or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the solution runs in linear time with clear grouping logic, such as reusing the encode_cyclic structure or calling encode_cyclic twice. It has no unnecessary complexity and no extra guards that change outputs. Any correct alternative approach gets full credit."}]}, "target_words": null, "reference": " return encode_cyclic(encode_cyclic(s))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/38", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef encode_cyclic(s: str):\n \"\"\"\n returns encoded string by cycling groups of three characters.\n \"\"\"\n # split string to groups. Each of length 3.\n groups = [s[(3 * i):min((3 * i + 3), len(s))] for i in range((len(s) + 2) // 3)]\n # cycle elements in each group. Unless group has fewer elements than 3.\n groups = [(group[1:] + group[0]) if len(group) == 3 else group for group in groups]\n return \"\".join(groups)\n\n\ndef decode_cyclic(s: str):\n \"\"\"\n takes as input string encoded with encode_cyclic function. Returns decoded string.\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n from random import randint, choice\n import string\n\n letters = string.ascii_lowercase\n for _ in range(100):\n str = ''.join(choice(letters) for i in range(randint(10, 20)))\n encoded_str = encode_cyclic(str)\n assert candidate(encoded_str) == str\n\n", "entry_point": "decode_cyclic"} {"task": "code_ru_6509ac27077f52a359df2dd2", "id": "6509ac27077f52a359df2dd2", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/106", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef f(n):\n \"\"\" Implement the function f that takes n as a parameter,\n and returns a list of size n, such that the value of the element at index i is the factorial of i if i is even\n or the sum of numbers from 1 to i otherwise.\n i starts from 1.\n the factorial of i is the multiplication of the numbers from 1 to i (1 * 2 * ... * i).\n Example:\n f(5) == [1, 2, 6, 24, 15]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on the 1-based indexing semantics: for each i from 1 to n, the element is i! when i is even and 1+2+...+i when i is odd, so f(5) returns exactly [1, 2, 6, 24, 15]. Mentally executing the body on this example must reproduce it. Off-by-one errors, such as starting i at 0 or swapping the even/odd rule, are major failures."}, {"id": "c2", "description": "Boundary handling: f(0) returns [] and f(1) returns [1] without crashing. For negative n, the body returns an empty list or behaves like range-based iteration. It must not loop forever or raise. A crash or non-termination on n=0 or n=1 caps this criterion at 6 or below."}, {"id": "c3", "description": "Validity of the returned body: it contains only the function body, with consistent indentation suitable for placement under the signature. It omits the def line, the docstring, and Markdown fences. It is syntactically valid Python, and the first line's indentation does not break the body. Any such defect lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: values are computed reasonably, for example with a running factorial or math.factorial and the arithmetic sum, and there are no wasteful or error-prone constructs. Correct alternative approaches get full credit. Extra guards that change outputs are not rewarded, and neither are comments."}]}, "target_words": null, "reference": " ret = []\n for i in range(1,n+1):\n if i%2 == 0:\n x = 1\n for j in range(1,i+1): x *= j\n ret += [x]\n else:\n x = 0\n for j in range(1,i+1): x += j\n ret += [x]\n return ret\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/106", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef f(n):\n \"\"\" Implement the function f that takes n as a parameter,\n and returns a list of size n, such that the value of the element at index i is the factorial of i if i is even\n or the sum of numbers from 1 to i otherwise.\n i starts from 1.\n the factorial of i is the multiplication of the numbers from 1 to i (1 * 2 * ... * i).\n Example:\n f(5) == [1, 2, 6, 24, 15]\n \"\"\"\n", "tests": "def check(candidate):\n\n assert candidate(5) == [1, 2, 6, 24, 15]\n assert candidate(7) == [1, 2, 6, 24, 15, 720, 28]\n assert candidate(1) == [1]\n assert candidate(3) == [1, 2, 6]\n", "entry_point": "f"} {"task": "code_ru_676eae37985451bd972e7679", "id": "676eae37985451bd972e7679", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/95", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef check_dict_case(dict):\n \"\"\"\n Given a dictionary, return True if all keys are strings in lower \n case or all keys are strings in upper case, else return False.\n The function should return False is the given dictionary is empty.\n Examples:\n check_dict_case({\"a\":\"apple\", \"b\":\"banana\"}) should return True.\n check_dict_case({\"a\":\"apple\", \"A\":\"banana\", \"B\":\"banana\"}) should return False.\n check_dict_case({\"a\":\"apple\", 8:\"banana\", \"a\":\"apple\"}) should return False.\n check_dict_case({\"Name\":\"John\", \"Age\":\"36\", \"City\":\"Houston\"}) should return False.\n check_dict_case({\"STATE\":\"NC\", \"ZIP\":\"12345\" }) should return True.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns False for an empty dict and for any dict containing a non-string key (e.g. {\"a\":1, 8:2}) without crashing, i.e. does not call .islower()/.isupper() on non-string keys."}, {"id": "c2", "description": "Returns True when all keys are lowercase strings or all are uppercase strings, and False for mixed case across keys ({\"a\":1,\"A\":2}) or within a key (\"Name\"), matching all docstring examples."}, {"id": "c3", "description": "Handles boundary keys consistently with str.islower/isupper semantics: keys with digits/underscores plus letters (e.g. \"zip_1\") are classified by their letters, and keys with no cased letters (e.g. \"123\" or \"\") cause False rather than being counted as both cases."}, {"id": "c4", "description": "Returned body is syntactically valid with consistent indentation suitable to place directly under the signature (no repeated def, docstring, or Markdown), and uses a simple linear pass over the keys."}]}, "target_words": null, "reference": " if len(dict.keys()) == 0:\n return False\n else:\n state = \"start\"\n for key in dict.keys():\n\n if isinstance(key, str) == False:\n state = \"mixed\"\n break\n if state == \"start\":\n if key.isupper():\n state = \"upper\"\n elif key.islower():\n state = \"lower\"\n else:\n break\n elif (state == \"upper\" and not key.isupper()) or (state == \"lower\" and not key.islower()):\n state = \"mixed\"\n break\n else:\n break\n return state == \"upper\" or state == \"lower\" \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/95", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef check_dict_case(dict):\n \"\"\"\n Given a dictionary, return True if all keys are strings in lower \n case or all keys are strings in upper case, else return False.\n The function should return False is the given dictionary is empty.\n Examples:\n check_dict_case({\"a\":\"apple\", \"b\":\"banana\"}) should return True.\n check_dict_case({\"a\":\"apple\", \"A\":\"banana\", \"B\":\"banana\"}) should return False.\n check_dict_case({\"a\":\"apple\", 8:\"banana\", \"a\":\"apple\"}) should return False.\n check_dict_case({\"Name\":\"John\", \"Age\":\"36\", \"City\":\"Houston\"}) should return False.\n check_dict_case({\"STATE\":\"NC\", \"ZIP\":\"12345\" }) should return True.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate({\"p\":\"pineapple\", \"b\":\"banana\"}) == True, \"First test error: \" + str(candidate({\"p\":\"pineapple\", \"b\":\"banana\"}))\n assert candidate({\"p\":\"pineapple\", \"A\":\"banana\", \"B\":\"banana\"}) == False, \"Second test error: \" + str(candidate({\"p\":\"pineapple\", \"A\":\"banana\", \"B\":\"banana\"}))\n assert candidate({\"p\":\"pineapple\", 5:\"banana\", \"a\":\"apple\"}) == False, \"Third test error: \" + str(candidate({\"p\":\"pineapple\", 5:\"banana\", \"a\":\"apple\"}))\n assert candidate({\"Name\":\"John\", \"Age\":\"36\", \"City\":\"Houston\"}) == False, \"Fourth test error: \" + str(candidate({\"Name\":\"John\", \"Age\":\"36\", \"City\":\"Houston\"}))\n assert candidate({\"STATE\":\"NC\", \"ZIP\":\"12345\" }) == True, \"Fifth test error: \" + str(candidate({\"STATE\":\"NC\", \"ZIP\":\"12345\" })) \n assert candidate({\"fruit\":\"Orange\", \"taste\":\"Sweet\" }) == True, \"Fourth test error: \" + str(candidate({\"fruit\":\"Orange\", \"taste\":\"Sweet\" })) \n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate({}) == False, \"1st edge test error: \" + str(candidate({}))\n\n", "entry_point": "check_dict_case"} {"task": "code_ru_69f9bf5fa93bd940e8e05780", "id": "69f9bf5fa93bd940e8e05780", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/124", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef valid_date(date):\n \"\"\"You have to write a function which validates a given date string and\n returns True if the date is valid otherwise False.\n The date is valid if all of the following rules are satisfied:\n 1. The date string is not empty.\n 2. The number of days is not less than 1 or higher than 31 days for months 1,3,5,7,8,10,12. And the number of days is not less than 1 or higher than 30 days for months 4,6,9,11. And, the number of days is not less than 1 or higher than 29 for the month 2.\n 3. The months should not be less than 1 or higher than 12.\n 4. The date should be in the format: mm-dd-yyyy\n\n for example: \n valid_date('03-11-2000') => True\n\n valid_date('15-01-2012') => False\n\n valid_date('04-0-2040') => False\n\n valid_date('06-04-2020') => True\n\n valid_date('06/04/2020') => False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly validates the mm-dd-yyyy format: returns True for '03-11-2000' and '06-04-2020', False for '06/04/2020', the empty string, and strings that cannot be split into three numeric parts. Malformed input such as '', 'abc', '01-02', or non-digit parts must yield False without raising an exception. A crash on malformed input is severe (score 6 or below)."}, {"id": "c2", "description": "Enforces month bounds 1–12 and the day bounds for each month: 1–31 for months 1,3,5,7,8,10,12; 1–30 for months 4,6,9,11; 1–29 for month 2. This means '15-01-2012' and '04-0-2040' return False, '02-29-2021' returns True (no leap-year check per spec), '04-31-2020' returns False, and '02-30-2020' returns False."}, {"id": "c3", "description": "The returned body is syntactically valid Python, consistently indented as a function body (the first line is not misindented), and contains no signature, docstring repetition or Markdown fences. It always returns a boolean."}, {"id": "c4", "description": "The logic is readable and appropriately simple, for example a split or regex followed by clear range checks. Extra guards that change outputs relative to the spec are not rewarded. A correct alternative approach such as a regex or datetime is acceptable only if its results match the spec, so datetime-based leap-year rejection of '02-29-2021' counts as an error."}]}, "target_words": null, "reference": " try:\n date = date.strip()\n month, day, year = date.split('-')\n month, day, year = int(month), int(day), int(year)\n if month < 1 or month > 12:\n return False\n if month in [1,3,5,7,8,10,12] and day < 1 or day > 31:\n return False\n if month in [4,6,9,11] and day < 1 or day > 30:\n return False\n if month == 2 and day < 1 or day > 29:\n return False\n except:\n return False\n\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/124", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef valid_date(date):\n \"\"\"You have to write a function which validates a given date string and\n returns True if the date is valid otherwise False.\n The date is valid if all of the following rules are satisfied:\n 1. The date string is not empty.\n 2. The number of days is not less than 1 or higher than 31 days for months 1,3,5,7,8,10,12. And the number of days is not less than 1 or higher than 30 days for months 4,6,9,11. And, the number of days is not less than 1 or higher than 29 for the month 2.\n 3. The months should not be less than 1 or higher than 12.\n 4. The date should be in the format: mm-dd-yyyy\n\n for example: \n valid_date('03-11-2000') => True\n\n valid_date('15-01-2012') => False\n\n valid_date('04-0-2040') => False\n\n valid_date('06-04-2020') => True\n\n valid_date('06/04/2020') => False\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate('03-11-2000') == True\n\n assert candidate('15-01-2012') == False\n\n assert candidate('04-0-2040') == False\n\n assert candidate('06-04-2020') == True\n\n assert candidate('01-01-2007') == True\n\n assert candidate('03-32-2011') == False\n\n assert candidate('') == False\n\n assert candidate('04-31-3000') == False\n\n assert candidate('06-06-2005') == True\n\n assert candidate('21-31-2000') == False\n\n assert candidate('04-12-2003') == True\n\n assert candidate('04122003') == False\n\n assert candidate('20030412') == False\n\n assert candidate('2003-04') == False\n\n assert candidate('2003-04-12') == False\n\n assert candidate('04-2003') == False\n", "entry_point": "valid_date"} {"task": "code_ru_6e25af64c5ac8b45b9a5ad1c", "id": "6e25af64c5ac8b45b9a5ad1c", "domain": "code", "language": "ru", "split": "validation", "group": "ru:HumanEval/62", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef derivative(xs: list):\n \"\"\" xs represent coefficients of a polynomial.\n xs[0] + xs[1] * x + xs[2] * x^2 + ....\n Return derivative of this polynomial in the same form.\n >>> derivative([3, 1, 2, 4, 5])\n [1, 4, 12, 20]\n >>> derivative([1, 2, 3])\n [2, 6]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on examples: the returned body computes [i * xs[i] for i in range(1, len(xs))], producing [1, 4, 12, 20] for [3, 1, 2, 4, 5] and [2, 6] for [1, 2, 3]. Coefficient order is preserved, so the result is not reversed and the constant term is dropped."}, {"id": "c2", "description": "Boundary cases: an empty list and a single-element list (a constant polynomial) both return [] without crashing. Zero and negative coefficients are multiplied correctly, for example [0, -3, 0] gives [-3, 0]. Trailing zeros are kept and not stripped, since the specification does not require normalization. Crashing on an empty or single-element input scores 6 or below."}, {"id": "c3", "description": "Validity of the returned body: the output is only the function body, with no signature, docstring or Markdown fences. Indentation is consistent and parses as a valid body. The body uses an explicit return of a list, not a generator or None. A broken first-line indent lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: the solution is a single linear pass, such as a list comprehension or enumerate. It does not mutate the input xs in a way that changes the caller's list, and it adds no extra guards that alter outputs. Any correct alternative approach earns full credit."}]}, "target_words": null, "reference": " return [(i * x) for i, x in enumerate(xs)][1:]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/62", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef derivative(xs: list):\n \"\"\" xs represent coefficients of a polynomial.\n xs[0] + xs[1] * x + xs[2] * x^2 + ....\n Return derivative of this polynomial in the same form.\n >>> derivative([3, 1, 2, 4, 5])\n [1, 4, 12, 20]\n >>> derivative([1, 2, 3])\n [2, 6]\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([3, 1, 2, 4, 5]) == [1, 4, 12, 20]\n assert candidate([1, 2, 3]) == [2, 6]\n assert candidate([3, 2, 1]) == [2, 2]\n assert candidate([3, 2, 1, 0, 4]) == [2, 2, 0, 16]\n assert candidate([1]) == []\n\n", "entry_point": "derivative"} {"task": "code_ru_7203df8e4e8461fe20a37245", "id": "7203df8e4e8461fe20a37245", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/90", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef next_smallest(lst):\n \"\"\"\n You are given a list of integers.\n Write a function next_smallest() that returns the 2nd smallest element of the list.\n Return None if there is no such element.\n \n next_smallest([1, 2, 3, 4, 5]) == 2\n next_smallest([5, 1, 4, 3, 2]) == 2\n next_smallest([]) == None\n next_smallest([1, 1]) == None\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns the second smallest DISTINCT value: [1,2,3,4,5]->2, [5,1,4,3,2]->2, [1,1]->None, [1,1,2]->2, [2,2,1,1,3]->2, and handles negatives such as [-5,-5,-3]->-3. Returning a duplicate of the minimum (e.g. 1 for [1,1,2]) is a significant correctness failure."}, {"id": "c2", "description": "Handles boundary inputs without crashing: [] -> None, a single-element list -> None, and an all-equal list -> None. An IndexError or ValueError on these inputs counts as a crash on in-domain input and is severe (score 6 or below)."}, {"id": "c3", "description": "The returned text is only a syntactically valid function body with consistent indentation (no signature, no docstring, no Markdown fences) that executes when inserted under the given signature. The first line's indentation must not break the body. The code must not mutate the input list in a way that changes results."}, {"id": "c4", "description": "Uses an appropriate, readable approach, such as sorted(set(lst)) or a single pass tracking the two smallest distinct values, with O(n log n) or better complexity. Any correct alternative algorithm gets full credit. No needless complexity or output-changing extra guards."}]}, "target_words": null, "reference": " lst = sorted(set(lst))\n return None if len(lst) < 2 else lst[1]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/90", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef next_smallest(lst):\n \"\"\"\n You are given a list of integers.\n Write a function next_smallest() that returns the 2nd smallest element of the list.\n Return None if there is no such element.\n \n next_smallest([1, 2, 3, 4, 5]) == 2\n next_smallest([5, 1, 4, 3, 2]) == 2\n next_smallest([]) == None\n next_smallest([1, 1]) == None\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([1, 2, 3, 4, 5]) == 2\n assert candidate([5, 1, 4, 3, 2]) == 2\n assert candidate([]) == None\n assert candidate([1, 1]) == None\n assert candidate([1,1,1,1,0]) == 1\n assert candidate([1, 0**0]) == None\n assert candidate([-35, 34, 12, -45]) == -35\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "next_smallest"} {"task": "code_ru_72e7ea86e42d947404db4979", "id": "72e7ea86e42d947404db4979", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/27", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef flip_case(string: str) -> str:\n \"\"\" For a given string, flip lowercase characters to uppercase and uppercase to lowercase.\n >>> flip_case('Hello')\n 'hELLO'\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring example: mentally executing the body on 'Hello' returns exactly 'hELLO'. Each lowercase letter becomes uppercase, each uppercase letter becomes lowercase, and non-letter characters (digits, spaces, punctuation) are left unchanged. Using str.swapcase() or an equivalent per-character mapping is fully correct."}, {"id": "c2", "description": "Boundary cases: the empty string returns '' without crashing. Strings with only non-alphabetic characters are returned unchanged. Mixed strings such as 'aB1 !' give 'Ab1 !'. Long strings are handled in linear time. Any crash or non-termination on a valid str input scores 6 or below. Rare Unicode case quirks (for example 'ß') are a minor issue at most."}, {"id": "c3", "description": "Validity of the returned body: the output is only the function body, with no signature, no repeated docstring and no Markdown fences. It has consistent indentation (for example 4 spaces) that would work when placed under the given signature, and it is syntactically valid Python that returns a str. If the first-line indentation is broken or any syntax error is present, this criterion is lowered."}, {"id": "c4", "description": "Readability and suitability: the solution is concise and idiomatic, such as `return string.swapcase()` or a clear comprehension. It avoids needless complexity, and it adds no guards or transformations that change outputs. Commentary or extra length is not rewarded."}]}, "target_words": null, "reference": " return string.swapcase()\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/27", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef flip_case(string: str) -> str:\n \"\"\" For a given string, flip lowercase characters to uppercase and uppercase to lowercase.\n >>> flip_case('Hello')\n 'hELLO'\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('') == ''\n assert candidate('Hello!') == 'hELLO!'\n assert candidate('These violent delights have violent ends') == 'tHESE VIOLENT DELIGHTS HAVE VIOLENT ENDS'\n", "entry_point": "flip_case"} {"task": "code_ru_7709960582d916d5b3e67727", "id": "7709960582d916d5b3e67727", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/23", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef strlen(string: str) -> int:\n \"\"\" Return length of given string\n >>> strlen('')\n 0\n >>> strlen('abc')\n 3\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns the number of characters in the input string. It yields 0 for '' and 3 for 'abc', and also gives correct counts for whitespace-only strings, strings with spaces, newlines, and multi-byte Unicode characters such as Cyrillic or emoji, where len counts code points rather than bytes. A different correct approach, such as a manual counting loop, gets full credit. A crash or non-termination on any str input scores 6 or below. Wrong values on rare inputs are penalized moderately."}, {"id": "c2", "description": "Syntactic validity: the response is only a function body, with no repeated signature, no docstring, and no Markdown fences. It is consistently indented, typically 4 spaces, so it compiles when placed under the given signature. If the first line's indentation breaks the body, or there are syntax errors or extraneous text, this criterion is lowered."}, {"id": "c3", "description": "Algorithmic suitability: the solution runs in O(1) using len, or at worst O(n), with no unnecessary conversions such as encoding to bytes, which would change the result for non-ASCII input. It adds no extra guards that alter outputs, for example special-casing None or stripping whitespace."}, {"id": "c4", "description": "Readability: the body is concise and idiomatic, ideally a single `return len(string)` or an equally clear equivalent. It has no superfluous commentary or complexity. Brevity alone is not rewarded beyond clarity."}]}, "target_words": null, "reference": " return len(string)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/23", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef strlen(string: str) -> int:\n \"\"\" Return length of given string\n >>> strlen('')\n 0\n >>> strlen('abc')\n 3\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('') == 0\n assert candidate('x') == 1\n assert candidate('asdasnakj') == 9\n", "entry_point": "strlen"} {"task": "code_ru_7cbabd04204092b06a874926", "id": "7cbabd04204092b06a874926", "domain": "code", "language": "ru", "split": "train", "group": "ru:HumanEval/110", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef exchange(lst1, lst2):\n \"\"\"In this problem, you will implement a function that takes two lists of numbers,\n and determines whether it is possible to perform an exchange of elements\n between them to make lst1 a list of only even numbers.\n There is no limit on the number of exchanged elements between lst1 and lst2.\n If it is possible to exchange elements between the lst1 and lst2 to make\n all the elements of lst1 to be even, return \"YES\".\n Otherwise, return \"NO\".\n For example:\n exchange([1, 2, 3, 4], [1, 2, 3, 4]) => \"YES\"\n exchange([1, 2, 3, 4], [1, 5, 3, 4]) => \"NO\"\n It is assumed that the input lists will be non-empty.\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Core correctness: the body returns \"YES\" exactly when the count of odd numbers in lst1 is less than or equal to the count of even numbers in lst2, and \"NO\" otherwise. It should reproduce the docstring examples: [1,2,3,4],[1,2,3,4] gives \"YES\" (2 odds vs 2 evens), and [1,2,3,4],[1,5,3,4] gives \"NO\" (2 odds vs 1 even). Any correct equivalent algorithm earns full credit."}, {"id": "c2", "description": "Boundary cases: the code handles these correctly without crashing. lst1 already all even gives \"YES\" even when lst2 has no evens. Negative numbers have their parity judged correctly; Python's % 2 handles negative odds such as -3 correctly. Zero counts as even. Large values work, and lists of different lengths work. Non-termination or a crash on valid non-empty integer lists scores 6 or below. Wrong output on a rare valid input is penalized moderately."}, {"id": "c3", "description": "Output format and validity: the result is only a function body, with no signature, no docstring and no Markdown. The indentation is consistent and syntactically valid as the body of the def. The return values are exactly the strings \"YES\" and \"NO\", with the right case and no extra text. A broken indentation on the first line or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach runs in linear time using parity counting, with no exponential search over swaps and no mutation of the input lists that would affect the result. The code is concise and clear. Extra commentary or guards that change outputs are not rewarded."}]}, "target_words": null, "reference": " odd = 0\n even = 0\n for i in lst1:\n if i%2 == 1:\n odd += 1\n for i in lst2:\n if i%2 == 0:\n even += 1\n if even >= odd:\n return \"YES\"\n return \"NO\"\n \n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/110", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef exchange(lst1, lst2):\n \"\"\"In this problem, you will implement a function that takes two lists of numbers,\n and determines whether it is possible to perform an exchange of elements\n between them to make lst1 a list of only even numbers.\n There is no limit on the number of exchanged elements between lst1 and lst2.\n If it is possible to exchange elements between the lst1 and lst2 to make\n all the elements of lst1 to be even, return \"YES\".\n Otherwise, return \"NO\".\n For example:\n exchange([1, 2, 3, 4], [1, 2, 3, 4]) => \"YES\"\n exchange([1, 2, 3, 4], [1, 5, 3, 4]) => \"NO\"\n It is assumed that the input lists will be non-empty.\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([1, 2, 3, 4], [1, 2, 3, 4]) == \"YES\"\n assert candidate([1, 2, 3, 4], [1, 5, 3, 4]) == \"NO\"\n assert candidate([1, 2, 3, 4], [2, 1, 4, 3]) == \"YES\" \n assert candidate([5, 7, 3], [2, 6, 4]) == \"YES\"\n assert candidate([5, 7, 3], [2, 6, 3]) == \"NO\" \n assert candidate([3, 2, 6, 1, 8, 9], [3, 5, 5, 1, 1, 1]) == \"NO\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([100, 200], [200, 200]) == \"YES\"\n\n", "entry_point": "exchange"} {"task": "code_ru_8182fa2bcb9b8dc3066da79c", "id": "8182fa2bcb9b8dc3066da79c", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/58", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef common(l1: list, l2: list):\n \"\"\"Return sorted unique common elements for two lists.\n >>> common([1, 4, 3, 34, 653, 2, 5], [5, 7, 1, 5, 9, 653, 121])\n [1, 5, 653]\n >>> common([5, 3, 2, 8], [3, 2])\n [2, 3]\n\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body yields [1, 5, 653] for the first example and [2, 3] for the second. The output must be a list, not a set or a generator, and must be sorted in ascending order."}, {"id": "c2", "description": "Uniqueness and boundary cases: duplicates present in either list appear only once in the output. An empty l1, an empty l2, or two disjoint lists return []. Negative numbers and lists with all-identical elements are handled correctly. Crashing on any of these in-domain inputs scores 6 or below."}, {"id": "c3", "description": "Validity of the returned body: it is only a function body, with no signature, docstring, or Markdown fences. Indentation is consistent and syntactically valid when placed under the given signature, including the first line. It returns a value rather than printing it. Broken indentation or syntax lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: it uses a reasonable approach, such as sorted(set(l1) & set(l2)), or a correct equivalent such as nested loops with deduplication. It does not add extra guards that change outputs. Inputs are not mutated in a way that affects correctness. Any correct alternative algorithm receives full credit, and commentary or verbosity is not rewarded."}]}, "target_words": null, "reference": " ret = set()\n for e1 in l1:\n for e2 in l2:\n if e1 == e2:\n ret.add(e1)\n return sorted(list(ret))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/58", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef common(l1: list, l2: list):\n \"\"\"Return sorted unique common elements for two lists.\n >>> common([1, 4, 3, 34, 653, 2, 5], [5, 7, 1, 5, 9, 653, 121])\n [1, 5, 653]\n >>> common([5, 3, 2, 8], [3, 2])\n [2, 3]\n\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([1, 4, 3, 34, 653, 2, 5], [5, 7, 1, 5, 9, 653, 121]) == [1, 5, 653]\n assert candidate([5, 3, 2, 8], [3, 2]) == [2, 3]\n assert candidate([4, 3, 2, 8], [3, 2, 4]) == [2, 3, 4]\n assert candidate([4, 3, 2, 8], []) == []\n\n", "entry_point": "common"} {"task": "code_ru_958d560047c8215eaea13bc5", "id": "958d560047c8215eaea13bc5", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/118", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef get_closest_vowel(word):\n \"\"\"You are given a word. Your task is to find the closest vowel that stands between \n two consonants from the right side of the word (case sensitive).\n \n Vowels in the beginning and ending doesn't count. Return empty string if you didn't\n find any vowel met the above condition. \n\n You may assume that the given string contains English letter only.\n\n Example:\n get_closest_vowel(\"yogurt\") ==> \"u\"\n get_closest_vowel(\"FULL\") ==> \"U\"\n get_closest_vowel(\"quick\") ==> \"\"\n get_closest_vowel(\"ab\") ==> \"\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on core semantics: scans from the right and returns the first vowel (from aeiouAEIOU, case preserved) whose immediate left and right neighbors are both consonants (non-vowels). Excludes the first and last character positions. Produces \"u\" for \"yogurt\", \"U\" for \"FULL\", \"\" for \"quick\" and \"\" for \"ab\". Returns \"\" when no such vowel exists. Treats 'y' as a consonant. Wrong results on plausible words like \"easy\" (expected \"\") or \"bad\" (expected \"a\") count as significant errors."}, {"id": "c2", "description": "Boundary robustness: handles words of length 0, 1, 2 and 3 without IndexError and returns \"\" when length < 3. Never treats wrapped negative indices (e.g. word[-1] when checking index 0) as neighbors. Handles all-vowel and all-consonant words correctly. Adjacent vowels such as \"ali\" vs \"aei\" are judged by immediate neighbors only. Any crash or infinite loop on such inputs scores 6 or below."}, {"id": "c3", "description": "Valid returned body: only the function body with consistent indentation suitable for placement under the given signature, with no repeated def/docstring, no Markdown fences, and no syntax errors. A first line whose indentation breaks the body lowers this criterion."}, {"id": "c4", "description": "Algorithmic suitability and readability: a linear right-to-left scan or equivalent correct approach, with clear vowel-set definition and loop bounds. No extra guards that alter outputs; a different correct algorithm gets full credit, and commentary is not rewarded."}]}, "target_words": null, "reference": " if len(word) < 3:\n return \"\"\n\n vowels = {\"a\", \"e\", \"i\", \"o\", \"u\", \"A\", \"E\", 'O', 'U', 'I'}\n for i in range(len(word)-2, 0, -1):\n if word[i] in vowels:\n if (word[i+1] not in vowels) and (word[i-1] not in vowels):\n return word[i]\n return \"\"\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/118", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef get_closest_vowel(word):\n \"\"\"You are given a word. Your task is to find the closest vowel that stands between \n two consonants from the right side of the word (case sensitive).\n \n Vowels in the beginning and ending doesn't count. Return empty string if you didn't\n find any vowel met the above condition. \n\n You may assume that the given string contains English letter only.\n\n Example:\n get_closest_vowel(\"yogurt\") ==> \"u\"\n get_closest_vowel(\"FULL\") ==> \"U\"\n get_closest_vowel(\"quick\") ==> \"\"\n get_closest_vowel(\"ab\") ==> \"\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"yogurt\") == \"u\"\n assert candidate(\"full\") == \"u\"\n assert candidate(\"easy\") == \"\"\n assert candidate(\"eAsy\") == \"\"\n assert candidate(\"ali\") == \"\"\n assert candidate(\"bad\") == \"a\"\n assert candidate(\"most\") == \"o\"\n assert candidate(\"ab\") == \"\"\n assert candidate(\"ba\") == \"\"\n assert candidate(\"quick\") == \"\"\n assert candidate(\"anime\") == \"i\"\n assert candidate(\"Asia\") == \"\"\n assert candidate(\"Above\") == \"o\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "get_closest_vowel"} {"task": "code_ru_9c5c81e5964ba56cef0e6d4a", "id": "9c5c81e5964ba56cef0e6d4a", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/91", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef is_bored(S):\n \"\"\"\n You'll be given a string of words, and your task is to count the number\n of boredoms. A boredom is a sentence that starts with the word \"I\".\n Sentences are delimited by '.', '?' or '!'.\n \n For example:\n >>> is_bored(\"Hello world\")\n 0\n >>> is_bored(\"The sky is blue. The sun is shining. I love this weather\")\n 1\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly splits the string into sentences on all three delimiters '.', '?' and '!' and counts sentences whose first word is exactly \"I\". Docstring examples return 0 and 1. Leading whitespace after a delimiter is handled, e.g. \"Hi. I am\" gives 1."}, {"id": "c2", "description": "Matches the whole word \"I\" rather than a prefix. Words like \"It\", \"In\" and \"Ice\" at sentence start are not counted, and an \"I\" in the middle of a sentence is not counted. A sentence that is exactly \"I\" (for example \"I!\" or \"Hello. I\") is counted, with no IndexError on empty fragments."}, {"id": "c3", "description": "Handles edge inputs without crashing: the empty string returns 0, consecutive delimiters such as \"..!\" are handled, trailing delimiters are fine, and whitespace-only sentences cause no error. A crash on any of these inputs scores this criterion at 6 or below."}, {"id": "c4", "description": "The returned body is syntactically valid with consistent indentation suitable for placement under the signature, includes no repeated signature, docstring or Markdown, uses a linear-time approach (for example re.split or char scanning), and is readable."}]}, "target_words": null, "reference": " import re\n sentences = re.split(r'[.?!]\\s*', S)\n return sum(sentence[0:2] == 'I ' for sentence in sentences)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/91", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef is_bored(S):\n \"\"\"\n You'll be given a string of words, and your task is to count the number\n of boredoms. A boredom is a sentence that starts with the word \"I\".\n Sentences are delimited by '.', '?' or '!'.\n \n For example:\n >>> is_bored(\"Hello world\")\n 0\n >>> is_bored(\"The sky is blue. The sun is shining. I love this weather\")\n 1\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"Hello world\") == 0, \"Test 1\"\n assert candidate(\"Is the sky blue?\") == 0, \"Test 2\"\n assert candidate(\"I love It !\") == 1, \"Test 3\"\n assert candidate(\"bIt\") == 0, \"Test 4\"\n assert candidate(\"I feel good today. I will be productive. will kill It\") == 2, \"Test 5\"\n assert candidate(\"You and I are going for a walk\") == 0, \"Test 6\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "is_bored"} {"task": "code_ru_9cbd96095557abcc4f9173ab", "id": "9cbd96095557abcc4f9173ab", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/82", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef prime_length(string):\n \"\"\"Write a function that takes a string and returns True if the string\n length is a prime number or False otherwise\n Examples\n prime_length('Hello') == True\n prime_length('abcdcba') == True\n prime_length('kittens') == True\n prime_length('orange') == False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: mentally executing the body gives True for 'Hello' (5), 'abcdcba' (7) and 'kittens' (7), and False for 'orange' (6). Any mismatch is a major failure."}, {"id": "c2", "description": "Boundary cases for length: the empty string (length 0) and single-character strings (length 1) return False, and length 2 returns True. Lengths 4, 9 and 25 return False, and 3 and 13 return True. Crashing or looping forever on '' or a one-character string is severe (6 or below). A wrong value for 0 or 1 is a significant correctness error because these inputs are plausible."}, {"id": "c3", "description": "Algorithmic suitability: primality is tested by trial division up to sqrt(n), or by an equivalent correct method, and the code terminates quickly for large string lengths. A correct O(n) loop to n is acceptable with a minor deduction. Off-by-one loop bounds that misclassify squares of primes, such as 4, 9 or 25, count as correctness errors."}, {"id": "c4", "description": "Validity and readability: the output is only the function body, with no signature, docstring or Markdown. It uses consistent indentation that is valid inside the def, including the first line, and is syntactically valid Python that returns a bool. The code is clear and concise. Extraneous commentary or print statements lower the score slightly; broken indentation lowers it substantially."}]}, "target_words": null, "reference": " l = len(string)\n if l == 0 or l == 1:\n return False\n for i in range(2, l):\n if l % i == 0:\n return False\n return True\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/82", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef prime_length(string):\n \"\"\"Write a function that takes a string and returns True if the string\n length is a prime number or False otherwise\n Examples\n prime_length('Hello') == True\n prime_length('abcdcba') == True\n prime_length('kittens') == True\n prime_length('orange') == False\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate('Hello') == True\n assert candidate('abcdcba') == True\n assert candidate('kittens') == True\n assert candidate('orange') == False\n assert candidate('wow') == True\n assert candidate('world') == True\n assert candidate('MadaM') == True\n assert candidate('Wow') == True\n assert candidate('') == False\n assert candidate('HI') == True\n assert candidate('go') == True\n assert candidate('gogo') == False\n assert candidate('aaaaaaaaaaaaaaa') == False\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate('Madam') == True\n assert candidate('M') == False\n assert candidate('0') == False\n\n", "entry_point": "prime_length"} {"task": "code_ru_a59e1484690bd31c13e22adb", "id": "a59e1484690bd31c13e22adb", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/1", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nfrom typing import List\n\n\ndef separate_paren_groups(paren_string: str) -> List[str]:\n \"\"\" Input to this function is a string containing multiple groups of nested parentheses. Your goal is to\n separate those group into separate strings and return the list of those.\n Separate groups are balanced (each open brace is properly closed) and not nested within each other\n Ignore any spaces in the input string.\n >>> separate_paren_groups('( ) (( )) (( )( ))')\n ['()', '(())', '(()())']\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on the docstring example: for input '( ) (( )) (( )( ))' the body returns exactly ['()', '(())', '(()())'], splitting at each point where the nesting depth returns to zero and collecting characters in order."}, {"id": "c2", "description": "Space handling and boundary cases: all spaces are ignored, including leading, trailing, repeated and inside-group spaces. An empty string or a whitespace-only string returns an empty list without crashing. A single group such as '((()))' returns a one-element list. Adjacent groups with no spaces, such as '()(())', are separated correctly. No empty strings appear in the output."}, {"id": "c3", "description": "Algorithmic suitability: the body uses a linear single pass with a depth counter or stack and does not add extra guards or validation that change outputs on valid balanced input. Any correct alternative algorithm gets full credit. Quadratic string rebuilding is acceptable but not ideal."}, {"id": "c4", "description": "Validity and readability: the returned text is only a function body that is syntactically valid, with consistent indentation so that it compiles when placed under the signature. It contains no repeated signature, docstring, or Markdown fences. Variable names are clear. Score this criterion lower if the first line's indentation breaks the body."}]}, "target_words": null, "reference": " result = []\n current_string = []\n current_depth = 0\n\n for c in paren_string:\n if c == '(':\n current_depth += 1\n current_string.append(c)\n elif c == ')':\n current_depth -= 1\n current_string.append(c)\n\n if current_depth == 0:\n result.append(''.join(current_string))\n current_string.clear()\n\n return result\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/1", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "from typing import List\n\n\ndef separate_paren_groups(paren_string: str) -> List[str]:\n \"\"\" Input to this function is a string containing multiple groups of nested parentheses. Your goal is to\n separate those group into separate strings and return the list of those.\n Separate groups are balanced (each open brace is properly closed) and not nested within each other\n Ignore any spaces in the input string.\n >>> separate_paren_groups('( ) (( )) (( )( ))')\n ['()', '(())', '(()())']\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('(()()) ((())) () ((())()())') == [\n '(()())', '((()))', '()', '((())()())'\n ]\n assert candidate('() (()) ((())) (((())))') == [\n '()', '(())', '((()))', '(((())))'\n ]\n assert candidate('(()(())((())))') == [\n '(()(())((())))'\n ]\n assert candidate('( ) (( )) (( )( ))') == ['()', '(())', '(()())']\n", "entry_point": "separate_paren_groups"} {"task": "code_ru_a6220ac480639172cef15c30", "id": "a6220ac480639172cef15c30", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/125", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef split_words(txt):\n '''\n Given a string of words, return a list of words split on whitespace, if no whitespaces exists in the text you\n should split on commas ',' if no commas exists you should return the number of lower-case letters with odd order in the\n alphabet, ord('a') = 0, ord('b') = 1, ... ord('z') = 25\n Examples\n split_words(\"Hello world!\") ➞ [\"Hello\", \"world!\"]\n split_words(\"Hello,world!\") ➞ [\"Hello\", \"world!\"]\n split_words(\"abcdef\") == 3 \n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Whitespace branch: when the input contains a space (HumanEval checks ' ' in txt), the body returns txt.split(), so \"Hello world!\" gives [\"Hello\", \"world!\"]. Splitting with split() and no argument, which drops empty strings, is acceptable."}, {"id": "c2", "description": "Comma branch: when there is no whitespace but a comma is present, the body splits on commas, so \"Hello,world!\" gives [\"Hello\", \"world!\"]. A common accepted form is txt.replace(',', ' ').split(). Whitespace takes priority over commas: \"Hello world,!\" splits on whitespace."}, {"id": "c3", "description": "Count branch: when there is neither whitespace nor a comma, the body returns an int counting characters that are lowercase a–z and have an odd alphabet index (b, d, f, …). \"abcdef\" gives 3, \"aaabb\" gives 2, and the empty string gives 0 without crashing. Uppercase letters and non-letters are not counted. Treating other lowercase characters, such as non-ASCII ones, as countable is wrong. If this check fails, score it 6 or below."}, {"id": "c4", "description": "The output is only a function body with valid, consistent indentation, with no signature, no docstring and no Markdown. It runs without a syntax or indentation error and terminates on all string inputs."}]}, "target_words": null, "reference": " if \" \" in txt:\n return txt.split()\n elif \",\" in txt:\n return txt.replace(',',' ').split()\n else:\n return len([i for i in txt if i.islower() and ord(i)%2 == 0])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/125", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef split_words(txt):\n '''\n Given a string of words, return a list of words split on whitespace, if no whitespaces exists in the text you\n should split on commas ',' if no commas exists you should return the number of lower-case letters with odd order in the\n alphabet, ord('a') = 0, ord('b') = 1, ... ord('z') = 25\n Examples\n split_words(\"Hello world!\") ➞ [\"Hello\", \"world!\"]\n split_words(\"Hello,world!\") ➞ [\"Hello\", \"world!\"]\n split_words(\"abcdef\") == 3 \n '''\n", "tests": "def check(candidate):\n\n assert candidate(\"Hello world!\") == [\"Hello\",\"world!\"]\n assert candidate(\"Hello,world!\") == [\"Hello\",\"world!\"]\n assert candidate(\"Hello world,!\") == [\"Hello\",\"world,!\"]\n assert candidate(\"Hello,Hello,world !\") == [\"Hello,Hello,world\",\"!\"]\n assert candidate(\"abcdef\") == 3\n assert candidate(\"aaabb\") == 2\n assert candidate(\"aaaBb\") == 1\n assert candidate(\"\") == 0\n", "entry_point": "split_words"} {"task": "code_ru_ad36a41b726c08ee8e9cdc72", "id": "ad36a41b726c08ee8e9cdc72", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/103", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef rounded_avg(n, m):\n \"\"\"You are given two positive integers n and m, and your task is to compute the\n average of the integers from n through m (including n and m). \n Round the answer to the nearest integer and convert that to binary.\n If n is greater than m, return -1.\n Example:\n rounded_avg(1, 5) => \"0b11\"\n rounded_avg(7, 5) => -1\n rounded_avg(10, 20) => \"0b1111\"\n rounded_avg(20, 33) => \"0b11010\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns -1 (integer) when n > m, and for n <= m returns bin() of the rounded average of n..m as a '0b...' string, matching all docstring examples: (1,5)->'0b11', (7,5)->-1, (10,20)->'0b1111', (20,33)->'0b11010'."}, {"id": "c2", "description": "Rounding of .5 averages is consistent with the reference examples. The (20,33) case has average 26.5, which must give 26 ('0b11010'). Either Python's round() (banker's rounding) or an equivalent is acceptable. Half-up rounding that yields 27 on this example is a wrong value. Handles n == m, returning bin(n), with no crash."}, {"id": "c3", "description": "Uses an algorithm suited to the task, such as (n+m)/2 or (n+m)//2 with correct tie handling. It must not crash or become impractically slow for large n and m, for example by summing a huge range when float precision or time would fail. Note that a float division like (n+m)/2 loses precision for very large values and can give a wrong rounding, which is a moderate issue."}, {"id": "c4", "description": "The returned body is valid Python with consistent indentation suitable for placement under the given signature. It contains no repeated signature, docstring, or Markdown fences, and is readable and concise."}]}, "target_words": null, "reference": " if m < n:\n return -1\n summation = 0\n for i in range(n, m+1):\n summation += i\n return bin(round(summation/(m - n + 1)))\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/103", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef rounded_avg(n, m):\n \"\"\"You are given two positive integers n and m, and your task is to compute the\n average of the integers from n through m (including n and m). \n Round the answer to the nearest integer and convert that to binary.\n If n is greater than m, return -1.\n Example:\n rounded_avg(1, 5) => \"0b11\"\n rounded_avg(7, 5) => -1\n rounded_avg(10, 20) => \"0b1111\"\n rounded_avg(20, 33) => \"0b11010\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(1, 5) == \"0b11\"\n assert candidate(7, 13) == \"0b1010\"\n assert candidate(964,977) == \"0b1111001010\"\n assert candidate(996,997) == \"0b1111100100\"\n assert candidate(560,851) == \"0b1011000010\"\n assert candidate(185,546) == \"0b101101110\"\n assert candidate(362,496) == \"0b110101101\"\n assert candidate(350,902) == \"0b1001110010\"\n assert candidate(197,233) == \"0b11010111\"\n\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(7, 5) == -1\n assert candidate(5, 1) == -1\n assert candidate(5, 5) == \"0b101\"\n\n", "entry_point": "rounded_avg"} {"task": "code_ru_ae320af80705db156a571add", "id": "ae320af80705db156a571add", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/72", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef will_it_fly(q,w):\n '''\n Write a function that returns True if the object q will fly, and False otherwise.\n The object q will fly if it's balanced (it is a palindromic list) and the sum of its elements is less than or equal the maximum possible weight w.\n\n Example:\n will_it_fly([1, 2], 5) ➞ False \n # 1+2 is less than the maximum possible weight, but it's unbalanced.\n\n will_it_fly([3, 2, 3], 1) ➞ False\n # it's balanced, but 3+2+3 is more than the maximum possible weight.\n\n will_it_fly([3, 2, 3], 9) ➞ True\n # 3+2+3 is less than the maximum possible weight, and it's balanced.\n\n will_it_fly([3], 5) ➞ True\n # 3 is less than the maximum possible weight, and it's balanced.\n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns True exactly when q is a palindrome (q equals its reverse) AND sum(q) <= w, and False otherwise. Mentally executing the body on all four docstring examples gives False, False, True, True respectively."}, {"id": "c2", "description": "Handles boundary cases correctly. The sum equal to w counts as flying (<=, not <). An empty list counts as balanced with sum 0, so it flies when w >= 0. Negative elements or a negative w give results consistent with the sum comparison. Single-element lists are palindromes. None of these inputs causes a crash or non-termination."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation suitable for placement directly under the given signature. It contains no repeated def line, docstring, or Markdown fences. A body whose first-line indentation breaks execution should score low here."}, {"id": "c4", "description": "Uses a suitable linear-time approach, such as comparing q to q[::-1] or using a two-pointer check together with sum. Avoids unnecessary complexity or extra guards that change outputs for valid inputs. The code is readable and concise."}]}, "target_words": null, "reference": " if sum(q) > w:\n return False\n\n i, j = 0, len(q)-1\n while i int:\n \"\"\" Find how many times a given substring can be found in the original string. Count overlaping cases.\n >>> how_many_times('', 'a')\n 0\n >>> how_many_times('aaa', 'a')\n 3\n >>> how_many_times('aaaa', 'aa')\n 3\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly counts overlapping occurrences: returns 0 for ('', 'a'), 3 for ('aaa', 'a'), 3 for ('aaaa', 'aa'), and handles a substring longer than the string (returns 0) without crashing. Using non-overlapping str.count is a correctness failure on the overlap examples."}, {"id": "c2", "description": "Handles boundary inputs without crashing or looping forever: an empty string, a substring equal to the whole string (returns 1), and a substring that never occurs (returns 0). For an empty substring, any terminating, defensible result is acceptable, since the spec does not define that case. A crash or infinite loop on the in-domain cases is severe and caps this criterion at 6."}, {"id": "c3", "description": "The returned body is syntactically valid Python. Its indentation is consistent and fits directly under the given signature. It contains only the body, with no repeated signature, no docstring and no Markdown fences. Broken first-line indentation or an included signature lowers this score."}, {"id": "c4", "description": "Uses a suitable algorithm: an O(n·m) sliding comparison, repeated find with a start offset advanced by 1, or an equivalent correct approach. The code is readable and has no needless complexity. Extra guards that change the specified outputs are penalized, and extra commentary earns no credit."}]}, "target_words": null, "reference": " times = 0\n\n for i in range(len(string) - len(substring) + 1):\n if string[i:i+len(substring)] == substring:\n times += 1\n\n return times\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/18", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef how_many_times(string: str, substring: str) -> int:\n \"\"\" Find how many times a given substring can be found in the original string. Count overlaping cases.\n >>> how_many_times('', 'a')\n 0\n >>> how_many_times('aaa', 'a')\n 3\n >>> how_many_times('aaaa', 'aa')\n 3\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate('', 'x') == 0\n assert candidate('xyxyxyx', 'x') == 4\n assert candidate('cacacacac', 'cac') == 4\n assert candidate('john doe', 'john') == 1\n", "entry_point": "how_many_times"} {"task": "code_ru_b4e0d5ad4ef83ab1ea80e083", "id": "b4e0d5ad4ef83ab1ea80e083", "domain": "code", "language": "ru", "split": "validation", "group": "ru:HumanEval/84", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef solve(N):\n \"\"\"Given a positive integer N, return the total sum of its digits in binary.\n \n Example\n For N = 1000, the sum of digits will be 1 the output should be \"1\".\n For N = 150, the sum of digits will be 6 the output should be \"110\".\n For N = 147, the sum of digits will be 12 the output should be \"1100\".\n \n Variables:\n @N integer\n Constraints: 0 ≤ N ≤ 10000.\n Output:\n a string of binary number\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body computes the sum of the decimal digits of N and returns that sum as a binary string with no '0b' prefix. Mentally executing it gives \"1\" for N=1000, \"110\" for N=150 and \"1100\" for N=147. A body that converts N itself to binary, or returns an int, scores low."}, {"id": "c2", "description": "Boundary cases: N=0 returns \"0\" (bin(0)[2:] gives '0'; format(0,'b') also gives '0'). Large values within 0..10000 such as 9999 (sum 36 -> \"100100\") and 10000 (\"1\") are handled. A crash or non-termination on any input in 0..10000 scores 6 or below. A wrong result for N=0 alone is a moderate error, since 0 is explicitly in the stated domain."}, {"id": "c3", "description": "Validity: the returned text is only a function body with consistent indentation suitable for placement under the signature. It contains no repeated def line, docstring or Markdown fences, and it is syntactically valid Python. A broken first-line indentation or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Algorithm and readability: the approach is simple and appropriate, for example summing digits via str(N) or a mod-10 loop, then using bin()[2:] or format(...,'b'). Any correct alternative earns full credit. Extra guards that change outputs are penalized, and unnecessary complexity or commentary is not rewarded."}]}, "target_words": null, "reference": " return bin(sum(int(i) for i in str(N)))[2:]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/84", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef solve(N):\n \"\"\"Given a positive integer N, return the total sum of its digits in binary.\n \n Example\n For N = 1000, the sum of digits will be 1 the output should be \"1\".\n For N = 150, the sum of digits will be 6 the output should be \"110\".\n For N = 147, the sum of digits will be 12 the output should be \"1100\".\n \n Variables:\n @N integer\n Constraints: 0 ≤ N ≤ 10000.\n Output:\n a string of binary number\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(1000) == \"1\", \"Error\"\n assert candidate(150) == \"110\", \"Error\"\n assert candidate(147) == \"1100\", \"Error\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate(333) == \"1001\", \"Error\"\n assert candidate(963) == \"10010\", \"Error\"\n\n", "entry_point": "solve"} {"task": "code_ru_b7183b5ec609e864e3741970", "id": "b7183b5ec609e864e3741970", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/3", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nfrom typing import List\n\n\ndef below_zero(operations: List[int]) -> bool:\n \"\"\" You're given a list of deposit and withdrawal operations on a bank account that starts with\n zero balance. Your task is to detect if at any point the balance of account fallls below zero, and\n at that point function should return True. Otherwise it should return False.\n >>> below_zero([1, 2, 3])\n False\n >>> below_zero([1, 2, -4, 5])\n True\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness of running balance logic: the body starts balance at 0, adds each operation in order, and returns True as soon as the balance becomes strictly less than zero (not <= 0); returns False after processing all operations otherwise. Docstring examples must yield False for [1, 2, 3] and True for [1, 2, -4, 5]."}, {"id": "c2", "description": "Boundary cases: empty list returns False; a list whose balance reaches exactly zero but never negative (e.g. [1, -1]) returns False; a first operation that is negative (e.g. [-1]) returns True; a dip below zero followed by recovery (e.g. [1, -2, 5]) still returns True. Any crash on these is severe (score 6 or below)."}, {"id": "c3", "description": "Validity of returned body: output is only the function body, without the signature, the docstring, or Markdown fences. It uses consistent indentation, including the first line, so it parses when placed under the given signature, and it introduces no undefined names."}, {"id": "c4", "description": "Algorithmic suitability and readability: uses a single O(n) pass (an early return or an equivalent such as checking the prefix sums is acceptable) with clear variable names. It must not add guards that change outputs, and it must not add unnecessary complexity."}]}, "target_words": null, "reference": " balance = 0\n\n for op in operations:\n balance += op\n if balance < 0:\n return True\n\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/3", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "from typing import List\n\n\ndef below_zero(operations: List[int]) -> bool:\n \"\"\" You're given a list of deposit and withdrawal operations on a bank account that starts with\n zero balance. Your task is to detect if at any point the balance of account fallls below zero, and\n at that point function should return True. Otherwise it should return False.\n >>> below_zero([1, 2, 3])\n False\n >>> below_zero([1, 2, -4, 5])\n True\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate([]) == False\n assert candidate([1, 2, -3, 1, 2, -3]) == False\n assert candidate([1, 2, -4, 5, 6]) == True\n assert candidate([1, -1, 2, -2, 5, -5, 4, -4]) == False\n assert candidate([1, -1, 2, -2, 5, -5, 4, -5]) == True\n assert candidate([1, -2, 2, -2, 5, -5, 4, -4]) == True\n", "entry_point": "below_zero"} {"task": "code_ru_b7deb9cca1f625e09628e8c2", "id": "b7deb9cca1f625e09628e8c2", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/61", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef correct_bracketing(brackets: str):\n \"\"\" brackets is a string of \"(\" and \")\".\n return True if every opening bracket has a corresponding closing bracket.\n\n >>> correct_bracketing(\"(\")\n False\n >>> correct_bracketing(\"()\")\n True\n >>> correct_bracketing(\"(()())\")\n True\n >>> correct_bracketing(\")(()\")\n False\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns True exactly when the string is a balanced bracket sequence. A running depth counter never goes negative and ends at zero. All four docstring examples produce the expected results, including \")(()\" -> False, where the count is balanced but the order is wrong."}, {"id": "c2", "description": "Boundary cases: the empty string returns True. A lone \")\" returns False. Strings that close too early, such as \"())(\", return False. Strings that leave brackets unclosed, such as \"((\", return False. Long inputs of any length terminate without crashing."}, {"id": "c3", "description": "Algorithmic suitability: uses an O(n) single pass with a counter or stack. Does not use repeated string replacement loops that are quadratic or may fail to terminate. Any different correct approach gets full credit."}, {"id": "c4", "description": "Validity and readability: the output is only a function body with consistent indentation, beginning with the correct indentation on the first line. It has no repeated signature, docstring, or Markdown fences, and it parses as valid Python. Variable names are clear, and the body contains no output-changing extra guards."}]}, "target_words": null, "reference": " depth = 0\n for b in brackets:\n if b == \"(\":\n depth += 1\n else:\n depth -= 1\n if depth < 0:\n return False\n return depth == 0\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/61", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef correct_bracketing(brackets: str):\n \"\"\" brackets is a string of \"(\" and \")\".\n return True if every opening bracket has a corresponding closing bracket.\n\n >>> correct_bracketing(\"(\")\n False\n >>> correct_bracketing(\"()\")\n True\n >>> correct_bracketing(\"(()())\")\n True\n >>> correct_bracketing(\")(()\")\n False\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate(\"()\")\n assert candidate(\"(()())\")\n assert candidate(\"()()(()())()\")\n assert candidate(\"()()((()()())())(()()(()))\")\n assert not candidate(\"((()())))\")\n assert not candidate(\")(()\")\n assert not candidate(\"(\")\n assert not candidate(\"((((\")\n assert not candidate(\")\")\n assert not candidate(\"(()\")\n assert not candidate(\"()()(()())())(()\")\n assert not candidate(\"()()(()())()))()\")\n\n", "entry_point": "correct_bracketing"} {"task": "code_ru_b91e8412a4a25aa0d259f700", "id": "b91e8412a4a25aa0d259f700", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/85", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef add(lst):\n \"\"\"Given a non-empty list of integers lst. add the even elements that are at odd indices..\n\n\n Examples:\n add([4, 2, 6, 7]) ==> 2 \n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns the sum of elements that are both at odd indices (1, 3, 5, ...) and even-valued (x % 2 == 0). For example, add([4, 2, 6, 7]) returns 2. Elements at even indices are excluded even when they are even, such as 4 and 6 in the example. Odd values at odd indices are excluded."}, {"id": "c2", "description": "Boundary cases: these inputs are handled without a crash and give the correct value. A single-element list returns 0. A list with no qualifying elements returns 0. Negative even numbers at odd indices are included, since in Python -4 % 2 == 0. Zero at an odd index contributes 0 without error. Mistaking index parity for value parity, or using 1-based indexing, is a severe correctness failure."}, {"id": "c3", "description": "Validity: the returned text is only a function body, with consistent indentation that is valid inside the def. It has no repeated signature, no docstring and no Markdown fences. It is syntactically valid Python and does not mutate the input."}, {"id": "c4", "description": "Algorithm and readability: the solution makes one linear pass, for example sum(lst[i] for i in range(1, len(lst), 2) if lst[i] % 2 == 0) or an equivalent clear loop. It has no unnecessary complexity and no extra guards that change outputs."}]}, "target_words": null, "reference": " return sum([lst[i] for i in range(1, len(lst), 2) if lst[i]%2 == 0])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/85", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef add(lst):\n \"\"\"Given a non-empty list of integers lst. add the even elements that are at odd indices..\n\n\n Examples:\n add([4, 2, 6, 7]) ==> 2 \n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([4, 88]) == 88\n assert candidate([4, 5, 6, 7, 2, 122]) == 122\n assert candidate([4, 0, 6, 7]) == 0\n assert candidate([4, 4, 6, 8]) == 12\n\n # Check some edge cases that are easy to work out by hand.\n \n", "entry_point": "add"} {"task": "code_ru_b91ff3bbee251c11a3970dde", "id": "b91ff3bbee251c11a3970dde", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/143", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef words_in_sentence(sentence):\n \"\"\"\n You are given a string representing a sentence,\n the sentence contains some words separated by a space,\n and you have to return a string that contains the words from the original sentence,\n whose lengths are prime numbers,\n the order of the words in the new string should be the same as the original one.\n\n Example 1:\n Input: sentence = \"This is a test\"\n Output: \"is\"\n\n Example 2:\n Input: sentence = \"lets go for swimming\"\n Output: \"go for\"\n\n Constraints:\n * 1 <= len(sentence) <= 100\n * sentence contains only letters\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correct primality filtering: the body keeps exactly the words whose length is prime (2, 3, 5, 7, ...) and excludes lengths 1, 4, 6, 8, 9, etc. The length-1 word \"a\" must be rejected. Mentally executing the body on \"This is a test\" yields \"is\", and on \"lets go for swimming\" yields \"go for\". Any misclassification of 1 or of small composites such as 4 or 9 is a significant correctness failure."}, {"id": "c2", "description": "Output formatting and order: the selected words are joined by single spaces in their original order, with no leading or trailing spaces. When no word qualifies (e.g., \"a\" or \"This\"), the result is the empty string \"\". A crash or non-termination on any valid sentence of length 1..100 caps this criterion at 6 or below."}, {"id": "c3", "description": "Syntactic validity: the returned text is only a function body. It has no repeated signature, docstring, or Markdown fences. Its indentation is consistent, so it compiles when placed under the def line. A first line with broken indentation or invalid syntax lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the prime check is correct and reasonably efficient for lengths ≤100, such as trial division or a small precomputed set. The code is clear and concise. No guards or special cases change outputs for valid inputs. Extra commentary or complexity earns no reward."}]}, "target_words": null, "reference": " new_lst = []\n for word in sentence.split():\n flg = 0\n if len(word) == 1:\n flg = 1\n for i in range(2, len(word)):\n if len(word)%i == 0:\n flg = 1\n if flg == 0 or len(word) == 2:\n new_lst.append(word)\n return \" \".join(new_lst)\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/143", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef words_in_sentence(sentence):\n \"\"\"\n You are given a string representing a sentence,\n the sentence contains some words separated by a space,\n and you have to return a string that contains the words from the original sentence,\n whose lengths are prime numbers,\n the order of the words in the new string should be the same as the original one.\n\n Example 1:\n Input: sentence = \"This is a test\"\n Output: \"is\"\n\n Example 2:\n Input: sentence = \"lets go for swimming\"\n Output: \"go for\"\n\n Constraints:\n * 1 <= len(sentence) <= 100\n * sentence contains only letters\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"This is a test\") == \"is\"\n assert candidate(\"lets go for swimming\") == \"go for\"\n assert candidate(\"there is no place available here\") == \"there is no place\"\n assert candidate(\"Hi I am Hussein\") == \"Hi am Hussein\"\n assert candidate(\"go for it\") == \"go for it\"\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate(\"here\") == \"\"\n assert candidate(\"here is\") == \"is\"\n\n", "entry_point": "words_in_sentence"} {"task": "code_ru_be33872b87f77f7042b6b609", "id": "be33872b87f77f7042b6b609", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/134", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef check_if_last_char_is_a_letter(txt):\n '''\n Create a function that returns True if the last character\n of a given string is an alphabetical character and is not\n a part of a word, and False otherwise.\n Note: \"word\" is a group of characters separated by space.\n\n Examples:\n check_if_last_char_is_a_letter(\"apple pie\") ➞ False\n check_if_last_char_is_a_letter(\"apple pi e\") ➞ True\n check_if_last_char_is_a_letter(\"apple pi e \") ➞ False\n check_if_last_char_is_a_letter(\"\") ➞ False \n '''\n", "rubric": {"criteria": [{"id": "c1", "description": "Core correctness: returns True exactly when the final character of txt is an alphabetic letter and either it is the whole string or the character before it is a space (i.e., the last space-separated token is a single letter). Must produce False for \"apple pie\", True for \"apple pi e\", False for \"apple pi e \", and False for \"\". Mentally execute these examples to confirm."}, {"id": "c2", "description": "Boundary cases handled without crashing: empty string returns False and raises no IndexError; a single letter such as \"a\" returns True; a single non-letter such as \"1\" or \" \" returns False; a trailing space returns False rather than inspecting an earlier token; a string like \"a b\" with multiple spaces before a final single letter returns True; and \"ab\" returns False. Indexing must be guarded so that txt[-1] and txt[-2] access are safe. A crash on empty or one-character input is severe and scores 6 or below."}, {"id": "c3", "description": "Validity of the returned body: the code is only the function body (no repeated signature or docstring, no Markdown fences), is consistently indented at function-body level including the first line, and is syntactically valid Python. Lower this criterion if the first line's indentation breaks the body."}, {"id": "c4", "description": "Algorithmic suitability and readability: uses a simple O(n) or O(1) approach, for example checking the last character plus its predecessor, or using split(' ')[-1] with len==1 and isalpha. Any correct alternative gets full credit. Logic is clear, and no extra guards change outputs relative to the specification."}]}, "target_words": null, "reference": " \n check = txt.split(' ')[-1]\n return True if len(check) == 1 and (97 <= ord(check.lower()) <= 122) else False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/134", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef check_if_last_char_is_a_letter(txt):\n '''\n Create a function that returns True if the last character\n of a given string is an alphabetical character and is not\n a part of a word, and False otherwise.\n Note: \"word\" is a group of characters separated by space.\n\n Examples:\n check_if_last_char_is_a_letter(\"apple pie\") ➞ False\n check_if_last_char_is_a_letter(\"apple pi e\") ➞ True\n check_if_last_char_is_a_letter(\"apple pi e \") ➞ False\n check_if_last_char_is_a_letter(\"\") ➞ False \n '''\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(\"apple\") == False\n assert candidate(\"apple pi e\") == True\n assert candidate(\"eeeee\") == False\n assert candidate(\"A\") == True\n assert candidate(\"Pumpkin pie \") == False\n assert candidate(\"Pumpkin pie 1\") == False\n assert candidate(\"\") == False\n assert candidate(\"eeeee e \") == False\n assert candidate(\"apple pie\") == False\n assert candidate(\"apple pi e \") == False\n\n # Check some edge cases that are easy to work out by hand.\n assert True\n\n", "entry_point": "check_if_last_char_is_a_letter"} {"task": "code_ru_c34a4b43c1489b8c710632ba", "id": "c34a4b43c1489b8c710632ba", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/2", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef truncate_number(number: float) -> float:\n \"\"\" Given a positive floating point number, it can be decomposed into\n and integer part (largest integer smaller than given number) and decimals\n (leftover part always smaller than 1).\n\n Return the decimal part of the number.\n >>> truncate_number(3.5)\n 0.5\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns the fractional part of a positive float, e.g. number % 1.0 or number - int(number) or number - math.floor(number), yielding 0.5 for 3.5. For positive inputs these approaches are equivalent and all deserve full credit. Penalize returning the integer part, rounding, or string-manipulation approaches that give wrong values (e.g. on scientific notation such as 1e-7 or large floats)."}, {"id": "c2", "description": "Boundary cases within the stated domain (positive floats): integer-valued inputs like 5.0 return 0.0; small values like 0.25 return 0.25; large values (e.g. 1e15+0.5) do not crash or loop. Any crash or non-termination on a positive float scores 6 or below. Behavior on 0 or negatives is not required by the spec; do not penalize either reasonable result, but penalize extra guards that change outputs for valid positive inputs."}, {"id": "c3", "description": "Format validity: output is only the function body, with consistent indentation that parses correctly when placed under the given signature. There is no repeated def line, no docstring, and no Markdown fences. Any needed import, such as math, appears inside the body. A broken first-line indentation or a syntax error lowers this score substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the solution uses a simple O(1) arithmetic expression and is concise and clear. Do not reward added commentary or unnecessary complexity."}]}, "target_words": null, "reference": " return number % 1.0\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/2", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef truncate_number(number: float) -> float:\n \"\"\" Given a positive floating point number, it can be decomposed into\n and integer part (largest integer smaller than given number) and decimals\n (leftover part always smaller than 1).\n\n Return the decimal part of the number.\n >>> truncate_number(3.5)\n 0.5\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate(3.5) == 0.5\n assert abs(candidate(1.33) - 0.33) < 1e-6\n assert abs(candidate(123.456) - 0.456) < 1e-6\n", "entry_point": "truncate_number"} {"task": "code_ru_da011f9fc58103c943af1ff5", "id": "da011f9fc58103c943af1ff5", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/96", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef count_up_to(n):\n \"\"\"Implement a function that takes an non-negative integer and returns an array of the first n\n integers that are prime numbers and less than n.\n for example:\n count_up_to(5) => [2,3]\n count_up_to(11) => [2,3,5,7]\n count_up_to(0) => []\n count_up_to(20) => [2,3,5,7,11,13,17,19]\n count_up_to(1) => []\n count_up_to(18) => [2,3,5,7,11,13,17]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: the body returns exactly the ascending list of all primes strictly less than n, e.g. count_up_to(5) -> [2,3], count_up_to(11) -> [2,3,5,7], count_up_to(20) -> [2,3,5,7,11,13,17,19], count_up_to(18) -> [2,3,5,7,11,13,17]. Excluding n itself when n is prime (e.g. 11 not included for n=11) is required; including it is a wrong value."}, {"id": "c2", "description": "Boundary cases in the non-negative domain: n=0, n=1, n=2 return [] and n=3 returns [2], with no crash or non-termination. Primality testing must not classify 0 or 1 as prime. A crash, an infinite loop or an IndexError on these inputs (e.g. a sieve allocating size n and indexing [1] when n=0) scores 6 or below."}, {"id": "c3", "description": "Algorithmic suitability: primality is determined correctly, either by trial division up to sqrt or full range, or by a Sieve of Eratosthenes. Performance is reasonable for moderately large n (e.g. 10^5) without pathological blowup. Any correct method earns full credit; added complexity or commentary earns no extra credit."}, {"id": "c4", "description": "Validity and readability: the output is only a function body, with no repeated signature, docstring or Markdown fences. It uses consistent indentation that parses correctly inside the def, including the first line. It returns a list and has clear variable names. Indentation or syntax errors that break the body lower this criterion substantially."}]}, "target_words": null, "reference": " primes = []\n for i in range(2, n):\n is_prime = True\n for j in range(2, i):\n if i % j == 0:\n is_prime = False\n break\n if is_prime:\n primes.append(i)\n return primes\n\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/96", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef count_up_to(n):\n \"\"\"Implement a function that takes an non-negative integer and returns an array of the first n\n integers that are prime numbers and less than n.\n for example:\n count_up_to(5) => [2,3]\n count_up_to(11) => [2,3,5,7]\n count_up_to(0) => []\n count_up_to(20) => [2,3,5,7,11,13,17,19]\n count_up_to(1) => []\n count_up_to(18) => [2,3,5,7,11,13,17]\n \"\"\"\n", "tests": "def check(candidate):\n\n assert candidate(5) == [2,3]\n assert candidate(6) == [2,3,5]\n assert candidate(7) == [2,3,5]\n assert candidate(10) == [2,3,5,7]\n assert candidate(0) == []\n assert candidate(22) == [2,3,5,7,11,13,17,19]\n assert candidate(1) == []\n assert candidate(18) == [2,3,5,7,11,13,17]\n assert candidate(47) == [2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43]\n assert candidate(101) == [2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, 37, 41, 43, 47, 53, 59, 61, 67, 71, 73, 79, 83, 89, 97]\n\n", "entry_point": "count_up_to"} {"task": "code_ru_dd4b826d079cf912cb50fbff", "id": "dd4b826d079cf912cb50fbff", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/73", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef smallest_change(arr):\n \"\"\"\n Given an array arr of integers, find the minimum number of elements that\n need to be changed to make the array palindromic. A palindromic array is an array that\n is read the same backwards and forwards. In one change, you can change one element to any other element.\n\n For example:\n smallest_change([1,2,3,5,4,7,9,6]) == 4\n smallest_change([1, 2, 3, 4, 3, 2, 2]) == 1\n smallest_change([1, 2, 3, 2, 1]) == 0\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body counts mismatched mirror pairs (arr[i] != arr[len(arr)-1-i] for i < len(arr)//2) and returns that count. It yields 4, 1 and 0 on the three docstring examples. Each mismatched pair counts once, not twice. The middle element of an odd-length array is never counted."}, {"id": "c2", "description": "Boundary cases: it returns 0 for an empty list and for a single-element list without crashing. It handles even and odd lengths, negative integers and repeated values correctly. It does not mutate the input in a way that changes the result."}, {"id": "c3", "description": "Validity and format: the response is only a function body, with no signature, docstring or Markdown fences. Indentation is consistent and syntactically valid when placed under the def line, including the first line. A body whose indentation breaks is scored 6 or below."}, {"id": "c4", "description": "Algorithmic suitability and readability: it uses an O(n) single pass, or an equivalent zip of the halves, with no unnecessary sorting, recursion or brute force. Variable names and structure are clear. A different algorithm that is correct gets full credit, and extra commentary earns no reward."}]}, "target_words": null, "reference": " ans = 0\n for i in range(len(arr) // 2):\n if arr[i] != arr[len(arr) - i - 1]:\n ans += 1\n return ans\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/73", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef smallest_change(arr):\n \"\"\"\n Given an array arr of integers, find the minimum number of elements that\n need to be changed to make the array palindromic. A palindromic array is an array that\n is read the same backwards and forwards. In one change, you can change one element to any other element.\n\n For example:\n smallest_change([1,2,3,5,4,7,9,6]) == 4\n smallest_change([1, 2, 3, 4, 3, 2, 2]) == 1\n smallest_change([1, 2, 3, 2, 1]) == 0\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate([1,2,3,5,4,7,9,6]) == 4\n assert candidate([1, 2, 3, 4, 3, 2, 2]) == 1\n assert candidate([1, 4, 2]) == 1\n assert candidate([1, 4, 4, 2]) == 1\n\n # Check some edge cases that are easy to work out by hand.\n assert candidate([1, 2, 3, 2, 1]) == 0\n assert candidate([3, 1, 1, 3]) == 0\n assert candidate([1]) == 0\n assert candidate([0, 1]) == 1\n\n", "entry_point": "smallest_change"} {"task": "code_ru_ec3d8503e9071009448a903b", "id": "ec3d8503e9071009448a903b", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/15", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef string_sequence(n: int) -> str:\n \"\"\" Return a string containing space-delimited numbers starting from 0 upto n inclusive.\n >>> string_sequence(0)\n '0'\n >>> string_sequence(5)\n '0 1 2 3 4 5'\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on docstring examples: string_sequence(0) returns exactly '0' and string_sequence(5) returns exactly '0 1 2 3 4 5'. Each number is separated by a single space, with no leading or trailing whitespace. The range includes n."}, {"id": "c2", "description": "Boundary behavior: n=0 yields '0'. A large n (e.g., 10**5) terminates efficiently in linear time, for example via ' '.join(map(str, range(n+1))). It must not use quadratic string concatenation that becomes impractically slow. Negative n is outside the stated domain, so do not penalize any non-crashing result such as ''. Do not reward extra guards that change outputs for valid n."}, {"id": "c3", "description": "Validity of the returned body: it is only the function body, with consistent indentation that parses correctly when placed under the signature. There is no repeated signature, docstring, or Markdown fences, and it contains a return statement producing a str. A broken first-line indentation or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "Readability and algorithmic suitability: the code is concise and idiomatic, such as a join over range(n+1), and is free of unnecessary complexity or commentary. Any correct alternative algorithm gets full credit."}]}, "target_words": null, "reference": " return ' '.join([str(x) for x in range(n + 1)])\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/15", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef string_sequence(n: int) -> str:\n \"\"\" Return a string containing space-delimited numbers starting from 0 upto n inclusive.\n >>> string_sequence(0)\n '0'\n >>> string_sequence(5)\n '0 1 2 3 4 5'\n \"\"\"\n", "tests": "\n\nMETADATA = {\n 'author': 'jt',\n 'dataset': 'test'\n}\n\n\ndef check(candidate):\n assert candidate(0) == '0'\n assert candidate(3) == '0 1 2 3'\n assert candidate(10) == '0 1 2 3 4 5 6 7 8 9 10'\n", "entry_point": "string_sequence"} {"task": "code_ru_eef25e9d9286f3a0a2c21153", "id": "eef25e9d9286f3a0a2c21153", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/101", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef words_string(s):\n \"\"\"\n You will be given a string of words separated by commas or spaces. Your task is\n to split the string into words and return an array of the words.\n \n For example:\n words_string(\"Hi, my name is John\") == [\"Hi\", \"my\", \"name\", \"is\", \"John\"]\n words_string(\"One, two, three, four, five, six\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctly splits on both commas and spaces: for \"Hi, my name is John\" returns [\"Hi\",\"my\",\"name\",\"is\",\"John\"] and for \"One, two, three, four, five, six\" returns the six words, with no empty strings produced by adjacent separators such as \", \"."}, {"id": "c2", "description": "Handles boundary inputs without crashing: an empty string returns [], and leading/trailing separators, repeated spaces or commas, and commas without following spaces (e.g. \"a,b\") yield only the non-empty words. A crash on empty input is severe (≤6). Stray empty strings on rarer inputs are moderate."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation suitable for placement under the given signature. It contains no repeated signature, docstring or Markdown fences, and it returns a list. Broken first-line indentation lowers this criterion."}, {"id": "c4", "description": "Uses a suitable linear-time approach, such as replacing commas with spaces and then calling split(), re.split with filtering, or a manual scan. The code is readable. Any correct algorithm gets full credit, and extra guards that alter outputs are not rewarded."}]}, "target_words": null, "reference": " if not s:\n return []\n\n s_list = []\n\n for letter in s:\n if letter == ',':\n s_list.append(' ')\n else:\n s_list.append(letter)\n\n s_list = \"\".join(s_list)\n return s_list.split()\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/101", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef words_string(s):\n \"\"\"\n You will be given a string of words separated by commas or spaces. Your task is\n to split the string into words and return an array of the words.\n \n For example:\n words_string(\"Hi, my name is John\") == [\"Hi\", \"my\", \"name\", \"is\", \"John\"]\n words_string(\"One, two, three, four, five, six\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert True, \"This prints if this assert fails 1 (good for debugging!)\"\n assert candidate(\"Hi, my name is John\") == [\"Hi\", \"my\", \"name\", \"is\", \"John\"]\n assert candidate(\"One, two, three, four, five, six\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n assert candidate(\"Hi, my name\") == [\"Hi\", \"my\", \"name\"]\n assert candidate(\"One,, two, three, four, five, six,\") == [\"One\", \"two\", \"three\", \"four\", \"five\", \"six\"]\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n assert candidate(\"\") == []\n assert candidate(\"ahmed , gamal\") == [\"ahmed\", \"gamal\"]\n\n", "entry_point": "words_string"} {"task": "code_ru_f109249d15e2707605dc9321", "id": "f109249d15e2707605dc9321", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/79", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\ndef decimal_to_binary(decimal):\n \"\"\"You will be given a number in decimal form and your task is to convert it to\n binary format. The function should return a string, with each character representing a binary\n number. Each character in the string will be '0' or '1'.\n\n There will be an extra couple of characters 'db' at the beginning and at the end of the string.\n The extra characters are there to help with the format.\n\n Examples:\n decimal_to_binary(15) # returns \"db1111db\"\n decimal_to_binary(32) # returns \"db100000db\"\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on examples: the function body returns exactly \"db1111db\" for 15 and \"db100000db\" for 32. The result is the binary representation of the input with no '0b' prefix and no leading zeros, wrapped in 'db' on both sides. Returning the string without the wrappers, or with a stray '0b', counts as a major failure."}, {"id": "c2", "description": "Boundary handling: decimal=0 returns \"db0db\", not \"dbdb\" and not an infinite loop. Large non-negative integers convert correctly and terminate. Any crash or non-termination on a non-negative integer input caps this criterion at 6 or below. Negative inputs are outside the stated domain, so behavior there should not be heavily penalized."}, {"id": "c3", "description": "Syntactic validity and format: the output is only a function body, with consistent indentation that is valid when placed under the signature. It contains no repeated def line, no docstring, and no Markdown fences. Indentation that breaks the body lowers this criterion substantially."}, {"id": "c4", "description": "Algorithmic suitability and readability: the approach is simple and efficient, such as bin(decimal)[2:] or a correct manual division loop. The code is clear and free of output-changing extra guards. Any correct algorithm earns full credit, and verbosity or commentary earns no extra credit."}]}, "target_words": null, "reference": " return \"db\" + bin(decimal)[2:] + \"db\"\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/79", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\ndef decimal_to_binary(decimal):\n \"\"\"You will be given a number in decimal form and your task is to convert it to\n binary format. The function should return a string, with each character representing a binary\n number. Each character in the string will be '0' or '1'.\n\n There will be an extra couple of characters 'db' at the beginning and at the end of the string.\n The extra characters are there to help with the format.\n\n Examples:\n decimal_to_binary(15) # returns \"db1111db\"\n decimal_to_binary(32) # returns \"db100000db\"\n \"\"\"\n", "tests": "def check(candidate):\n\n # Check some simple cases\n assert candidate(0) == \"db0db\"\n assert candidate(32) == \"db100000db\"\n assert candidate(103) == \"db1100111db\"\n assert candidate(15) == \"db1111db\", \"This prints if this assert fails 1 (good for debugging!)\"\n\n # Check some edge cases that are easy to work out by hand.\n assert True, \"This prints if this assert fails 2 (also good for debugging!)\"\n\n", "entry_point": "decimal_to_binary"} {"task": "code_ru_f3ed812930470f9a2bc1660f", "id": "f3ed812930470f9a2bc1660f", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/32", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\nimport math\n\n\ndef poly(xs: list, x: float):\n \"\"\"\n Evaluates polynomial with coefficients xs at point x.\n return xs[0] + xs[1] * x + xs[1] * x^2 + .... xs[n] * x^n\n \"\"\"\n return sum([coeff * math.pow(x, i) for i, coeff in enumerate(xs)])\n\n\ndef find_zero(xs: list):\n \"\"\" xs are coefficients of a polynomial.\n find_zero find x such that poly(x) = 0.\n find_zero returns only only zero point, even if there are many.\n Moreover, find_zero only takes list xs having even number of coefficients\n and largest non zero coefficient as it guarantees\n a solution.\n >>> round(find_zero([1, 2]), 2) # f(x) = 1 + 2x\n -0.5\n >>> round(find_zero([-6, 11, -6, 1]), 2) # (x - 1) * (x - 2) * (x - 3) = -6 + 11x - 6x^2 + x^3\n 1.0\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness on stated domain: for lists with an even number of coefficients and a nonzero last coefficient (odd-degree polynomial), the body returns an x with poly(xs, x) ≈ 0 within reasonable tolerance (e.g., |poly| < 1e-6 or round-to-2-decimals matches). Mentally verify [1,2] → -0.5 and [-6,11,-6,1] → a root (1.0 per doctest, though any true root should be accepted only if the rubric allows; doctest expects 1.0 when using standard bracketing from [-1,1] expansion)."}, {"id": "c2", "description": "Termination and robustness: the algorithm always terminates and does not crash for valid inputs, including roots far from the origin (e.g., [-1000000, 1]), large coefficients, roots exactly at bracket endpoints or at 0 (e.g., [0, 1]), and steep or flat polynomials. Bisection with bracket expansion is suitable. Newton's method without safeguards that can diverge, cycle, or divide by a zero derivative counts as severe (≤6). So does an infinite loop from a fixed bracket or an exact-equality stop condition."}, {"id": "c3", "description": "Algorithmic suitability: uses a sound root-finding approach such as bisection, which is guaranteed by the sign change of an odd-degree polynomial, a safeguarded secant/Newton method, or similar. It reaches adequate precision with a bounded iteration count. Uses poly() or an equivalent evaluation. Does not add guards or checks that change outputs for valid inputs."}, {"id": "c4", "description": "Syntactic validity and readability: the output is only the function body, with consistent indentation that is valid when placed under the signature and docstring. It contains no repeated signature and no Markdown fences, and has clear variable names. Lower this score if the first line's indentation breaks the body."}]}, "target_words": null, "reference": " begin, end = -1., 1.\n while poly(xs, begin) * poly(xs, end) > 0:\n begin *= 2.0\n end *= 2.0\n while end - begin > 1e-10:\n center = (begin + end) / 2.0\n if poly(xs, center) * poly(xs, begin) > 0:\n begin = center\n else:\n end = center\n return begin\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/32", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "import math\n\n\ndef poly(xs: list, x: float):\n \"\"\"\n Evaluates polynomial with coefficients xs at point x.\n return xs[0] + xs[1] * x + xs[1] * x^2 + .... xs[n] * x^n\n \"\"\"\n return sum([coeff * math.pow(x, i) for i, coeff in enumerate(xs)])\n\n\ndef find_zero(xs: list):\n \"\"\" xs are coefficients of a polynomial.\n find_zero find x such that poly(x) = 0.\n find_zero returns only only zero point, even if there are many.\n Moreover, find_zero only takes list xs having even number of coefficients\n and largest non zero coefficient as it guarantees\n a solution.\n >>> round(find_zero([1, 2]), 2) # f(x) = 1 + 2x\n -0.5\n >>> round(find_zero([-6, 11, -6, 1]), 2) # (x - 1) * (x - 2) * (x - 3) = -6 + 11x - 6x^2 + x^3\n 1.0\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n import math\n import random\n rng = random.Random(42)\n import copy\n for _ in range(100):\n ncoeff = 2 * rng.randint(1, 4)\n coeffs = []\n for _ in range(ncoeff):\n coeff = rng.randint(-10, 10)\n if coeff == 0:\n coeff = 1\n coeffs.append(coeff)\n solution = candidate(copy.deepcopy(coeffs))\n assert math.fabs(poly(coeffs, solution)) < 1e-4\n\n", "entry_point": "find_zero"} {"task": "code_ru_fa4098047365eb6b8cc46da6", "id": "fa4098047365eb6b8cc46da6", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/57", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef monotonic(l: list):\n \"\"\"Return True is list elements are monotonically increasing or decreasing.\n >>> monotonic([1, 2, 4, 20])\n True\n >>> monotonic([1, 20, 4, 10])\n False\n >>> monotonic([4, 1, 0, -10])\n True\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Returns True for non-strictly increasing or non-strictly decreasing lists and False otherwise, matching all three docstring examples ([1,2,4,20]→True, [1,20,4,10]→False, [4,1,0,-10]→True). Equal adjacent elements (e.g., [1,1,2], [3,3,3]) count as monotonic. Wrong results on common inputs are a major failure."}, {"id": "c2", "description": "Handles boundary inputs without crashing or looping forever: the empty list and single-element lists return True, as do all-equal lists and lists with negative values. A crash such as an IndexError on [] or [x] scores 6 or below. A wrong value on these rare inputs is a moderate deduction."}, {"id": "c3", "description": "The returned body is syntactically valid Python with consistent indentation suitable for placement directly under the signature. It does not repeat the def line or the docstring and contains no Markdown fences. Broken first-line indentation or a syntax error lowers this criterion substantially."}, {"id": "c4", "description": "The approach is algorithmically reasonable, such as a single pass, pairwise comparisons, or comparison against sorted(l) and sorted(l, reverse=True). It is readable and does not mutate the input list. Any correct algorithm earns full credit. Extra commentary or complexity earns nothing."}]}, "target_words": null, "reference": " if l == sorted(l) or l == sorted(l, reverse=True):\n return True\n return False\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/57", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef monotonic(l: list):\n \"\"\"Return True is list elements are monotonically increasing or decreasing.\n >>> monotonic([1, 2, 4, 20])\n True\n >>> monotonic([1, 20, 4, 10])\n False\n >>> monotonic([4, 1, 0, -10])\n True\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([1, 2, 4, 10]) == True\n assert candidate([1, 2, 4, 20]) == True\n assert candidate([1, 20, 4, 10]) == False\n assert candidate([4, 1, 0, -10]) == True\n assert candidate([4, 1, 1, 0]) == True\n assert candidate([1, 2, 3, 2, 5, 60]) == False\n assert candidate([1, 2, 3, 4, 5, 60]) == True\n assert candidate([9, 9, 9, 9]) == True\n\n", "entry_point": "monotonic"} {"task": "code_ru_fb65577a6e90df79401699d9", "id": "fb65577a6e90df79401699d9", "domain": "code", "language": "ru", "split": "test", "group": "ru:HumanEval/42", "prompt": "Реализуй функцию Python по приведённой спецификации. Верни ТОЛЬКО тело функции с правильными отступами, без повторения сигнатуры и docstring, без Markdown. Учитывай крайние случаи. Оригинальная спецификация ниже на английском; это намеренно не переведённый HumanEval, а русская инструкция к задаче.\n\n\n\ndef incr_list(l: list):\n \"\"\"Return list with elements incremented by 1.\n >>> incr_list([1, 2, 3])\n [2, 3, 4]\n >>> incr_list([5, 3, 5, 2, 3, 3, 9, 0, 123])\n [6, 4, 6, 3, 4, 4, 10, 1, 124]\n \"\"\"\n", "rubric": {"criteria": [{"id": "c1", "description": "Correctness: the body returns a new list in which every element equals the corresponding input element plus 1, with order and length preserved. Mentally executing it on [1, 2, 3] gives [2, 3, 4], and on [5, 3, 5, 2, 3, 3, 9, 0, 123] gives [6, 4, 6, 3, 4, 4, 10, 1, 124]. A crash or wrong output on these examples scores 6 or below."}, {"id": "c2", "description": "Boundary cases: an empty list returns []. Negative values (e.g. [-1, -5] gives [0, -4]), zero and very large integers are incremented correctly without error. Any exception or wrong result on an empty list, which is in the domain, scores 6 or below."}, {"id": "c3", "description": "Validity and format: the output is only a function body (no signature, docstring or Markdown fences). It has consistent indentation that compiles when placed under the given def line. It contains a return statement producing a list, not a generator, map object or None. Broken indentation on the first line lowers this score."}, {"id": "c4", "description": "Algorithm and readability: the body uses a linear, idiomatic approach, such as a list comprehension or a simple loop building a new list. It has no unnecessary complexity and no extra guards that alter outputs. Whether it mutates the input or builds a new list is acceptable either way, provided the returned value is correct."}]}, "target_words": null, "reference": " return [(e + 1) for e in l]\n", "license": "MIT", "source": {"url": "https://raw.githubusercontent.com/openai/human-eval/6d43fb980f9fee3c892a914eda09951f772ad10d/data/HumanEval.jsonl.gz", "revision": "6d43fb980f9fee3c892a914eda09951f772ad10d", "license": "MIT", "task_id": "HumanEval/42", "source_language": "en", "instruction_language": "ru", "reference_language": "python", "native_russian_description": false, "quality": "Human canonical solution; public-benchmark memorization unknown."}, "prefix": "\n\ndef incr_list(l: list):\n \"\"\"Return list with elements incremented by 1.\n >>> incr_list([1, 2, 3])\n [2, 3, 4]\n >>> incr_list([5, 3, 5, 2, 3, 3, 9, 0, 123])\n [6, 4, 6, 3, 4, 4, 10, 1, 124]\n \"\"\"\n", "tests": "\n\nMETADATA = {}\n\n\ndef check(candidate):\n assert candidate([]) == []\n assert candidate([3, 2, 1]) == [4, 3, 2]\n assert candidate([5, 2, 5, 2, 3, 3, 9, 0, 123]) == [6, 3, 6, 3, 4, 4, 10, 1, 124]\n\n", "entry_point": "incr_list"} {"task": "tweets_en_0a3a8a54bec7ea1f1fdb0633", "id": "0a3a8a54bec7ea1f1fdb0633", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463991704937168896", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 16 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\ni forgot how my grandma’s hand ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The judge must first quote the joined text at the seam (e.g., \"...my grandma's hand [continuation start]\"). The prefix ends on the noun \"hand\" followed by a space, so a natural continuation picks up as a verb phrase or clause about the hand (e.g., \"felt\", \"used to\", \"would\", \"smelled\") or a plural/possessive move that reads grammatically. Penalize re-typing \"hand\" or \"grandma's hand\", starting with a capitalized restart or sentence-initial hedge, or producing a joined sentence that is ungrammatical or doesn't finish the thought. A continuation that starts exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the lowercase, casual, personal and tender or nostalgic register of the prefix (\"i forgot how...\"). Fragments, run-ons, lowercase, \"..\", and raw emotional phrasing are fine. Heavily penalize assistant-style writing: hedges or bookish connectives (\"perhaps\", \"truly\", \"which reminds us\"), polished balanced clauses, and tidy morals or explanatory closers (e.g., \"cherish your loved ones while you can\"). A rough but heartfelt line beats a formal, sentimental-card rewrite."}, {"id": "c3", "description": "Substance: Reward a specific sensory or concrete detail about the grandma's hand, such as texture, a ring, lotion smell, a gesture like squeezing or tracing circles, or a remembered moment that triggered the recollection. Generic phrasing (\"felt so warm and loving\", \"meant everything\") with nothing distinctive caps naturalness around 8/10. Do not credit filler or padding clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation contains no links, no meta-commentary or explanation of the task, and no invented hashtags. The full tweet stays within 280 characters. A hashtag that fits the author's own playful voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 16, "reference": "felt like. I forgot how my uncle would slur when he speaks. I can hardly remember.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463991704937168896", "author_group_sha256": "b8643ea6f1400a9e40f7049a274eb271ceb1899e350293007136d5bb2d55ed01", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_0d00d3a69136d74147bd5a53", "id": "0d00d3a69136d74147bd5a53", "domain": "tweets", "language": "en", "split": "train", "group": "btc:463994133544071168", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 17 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nBoko Haram's ruthless, twisted leader: He's ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (\"...leader: He's \" + continuation's opening) before judging. The prefix ends after \"He's \" with a trailing space, so the continuation should begin with a predicate completing \"He's\" (e.g., an adjective, noun phrase, or verb-ing). Mark it down if it re-types \"He's\", restarts with a capitalized new sentence or hedge, or reads ungrammatically when joined. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the tweet's blunt, contemptuous, casual stance toward Boko Haram's leader (e.g., punchy fragments, informal phrasing, sharp scorn). Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, or explanatory/moralizing closers. A rough but lively jab outranks a hedged, formal rewrite, and softening or neutralizing the author's condemnation counts as a voice failure."}, {"id": "c3", "description": "Substance: Reward a specific image, concrete action, pointed detail, or actual biting joke about the leader (e.g., something he does, says, or resembles). Stock hyperbole or generic insults ('pure evil', 'the worst person ever') with nothing distinctive are weak; cap naturalness around 8 in that case. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation text—no links, meta-commentary, quotation marks around the answer, or explanation of the task—and the full tweet stays within 280 characters. A hashtag that sounds like the author's own voice is a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 17, "reference": "the sadistic architect of a campaign of mayhem and misery in Nigeria. But very little is ...", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463994133544071168", "author_group_sha256": "1df1b3e6e446ca90667251e5837269e51a6f27489fe7082679b1572a73fed219", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_151e47205bf4000f16414597", "id": "151e47205bf4000f16414597", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458974998678745088", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 17 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nDon't follow to unfollow cuz I ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with \"cuz I \" (a subject pronoun followed by a space), so the continuation should open with a verb or adverb that completes 'I ...', such as 'will notice', 'check', 'see you'. The judge should quote the joined text at the seam (e.g., 'cuz I [continuation start]') before deciding. Penalize a continuation that re-types 'I' or 'cuz', opens with a capitalized restart or a hedge like 'Perhaps', or fails to form a grammatical clause with 'I'. A continuation that starts exactly at the verb is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, slightly defiant social-media warning tone (the stance is against follow-for-unfollow behavior, written informally with 'cuz'). Slang, lowercase, nonstandard spelling, run-ons, fragments, and sassy threats or exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), tidy morals about online respect, balanced polished clauses, and explanatory or moralizing closers. A rough but lively line beats a formal one."}, {"id": "c3", "description": "Substance: Reward a concrete action, specific consequence, or actual joke tied to follow/unfollow behavior. Examples include saying the author checks who unfollowed, blocks people, or uses a named tracking habit, or adding a punchy comic threat. Stock hyperbole or generic filler like 'I don't like that stuff' is weak, and naturalness is capped around 8 when nothing distinctive is present. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation contains no links, meta-commentary, or task explanation, and the full tweet stays within 280 characters. Hashtags are discouraged by the assignment. If one sounds like the author's own playful voice, treat it as a minor flaw only, never one that outweighs voice or seam failures."}]}, "target_words": 17, "reference": "have & been having the app, so if u unfollow I'll unfollow u back jst like that.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458974998678745088", "author_group_sha256": "30aab9b497548234c32d1d523eab7a4f0613de2fc92fc37e1c7c01a0e922f97b", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_1b13bd4f5b42fb3770e9ca6c", "id": "1b13bd4f5b42fb3770e9ca6c", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458971198714839041", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 25 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI'm streaming War of the Vikings ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with a trailing space after the complete title 'War of the Vikings'. The judge should quote the joined text at the seam, e.g. '...War of the Vikings [continuation start]', before deciding. A good continuation picks up naturally from there, for example with 'right now', 'on twitch', 'tonight', 'come watch' or a lowercase clause. Penalize any of these: re-typing 'Vikings' or the title, a capitalized restart that reads like a new sentence ignoring the prefix, or a sentence-initial hedge. A continuation that simply starts where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps a casual streamer/gamer register, the kind of excited self-promotional or chill gamer chatter you would expect from someone announcing a stream. Informality, slang, lowercase, fragments, '..' run-ons and exclamations are fine. Heavily penalize assistant-style writing, including hedges or bookish connectives (perhaps, certainly, apparently), balanced polished clauses, tidy morals, and explanatory closers. A rough but lively line beats a formal, polished one."}, {"id": "c3", "description": "Substance: Reward concrete, specific content such as a named game element (axes, shield walls, duels, a class or map), a specific situation (getting wrecked, playing with friends, a first try), a concrete invite, or an actual joke. Score generic filler like 'it's so fun, come join the fun' or stock hyperbole as weak, and cap it around 8 when nothing distinctive accompanies it. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The continuation contains no URL or invented link, no meta-commentary or explanation of the task, and no quotation of the prefix. The full tweet stays within 280 characters. A playful hashtag that fits the author's voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 25, "reference": "at 5pm. If you have it, I won't be hiding the server so feel free to join in and try to cut my head off!", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458971198714839041", "author_group_sha256": "80eacb0bcdb6bf0e796e1ef57ef8897f3555d2f4c6e6abe15e04a1cee8b38486", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_1f2f976d3aa07470207924e3", "id": "1f2f976d3aa07470207924e3", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458971660297596929", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 16 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nBut i dont want to be ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with \"be \" (a trailing space after a verb needing a complement). Quote the joined text at the seam (e.g., \"...dont want to be [continuation]\") before judging. The continuation must supply a complement (adjective, noun phrase, location, state) that reads as a grammatical, natural completion of \"I dont want to be...\". It must not re-type \"be\" or other prefix words, and must not open with a capitalized restart or sentence-initial hedge. A continuation that simply starts where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the reluctant, whiny or resistant casual tone of \"But i dont want to be\" (lowercase, missing apostrophe, informal). Slang, lowercase, fragments, '..' run-ons and nonstandard spelling are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or self-help closers. A rough but lively complaint beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete reason, image, named thing, situation or real joke behind the reluctance (e.g., a particular place, person, chore or feeling). Generic phrasing or stock hyperbole with nothing distinctive caps naturalness around 8/10. Give no credit for filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check that the output has no links, no meta-commentary or task explanation, and that the full tweet stays within 280 characters. A hashtag is a minor flaw only if it sounds like the author's playful voice, and it must never outweigh a failure of seam or voice."}]}, "target_words": 16, "reference": "like that. Lebih rela cmtok dr dpt pgganti yg belom tntu baik and dpt trimak seadanya.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458971660297596929", "author_group_sha256": "13d69ca55a14e76273261e142d4bbd0736707547914c2f149aa5efe4e6839473", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_23fa7c24edfc1de81d910376", "id": "23fa7c24edfc1de81d910376", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439816227720937473", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 19 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\n♪All the kids have always known, ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends after a comma following 'All the kids have always known,' (song-lyric framing with ♪). The judge must quote the joined text at the seam (e.g., 'known, [first words]') before deciding. The continuation should flow as the next clause of the line or tweet, in lowercase or natural lyric continuation. It should not re-type 'known' or any prefix words, and it should not open with a capitalized restart or a sentence-initial hedge. A continuation that starts exactly after the comma is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, playful, sing-along tweet register implied by the ♪ lyric opener. Lowercase, slang, fragments, run-ons, exclamations, or a closing ♪ are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, or explanatory or moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action, or actual joke. Examples include a lyric-like twist, a recognizable cultural reference, or a personal punchline tying the song line to the author's life. Stock hyperbole or generic filler ('life is crazy', 'so true') counts as weak, and naturalness is capped around 8 if nothing distinctive appears. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene, a minor weight below c1–c3: Check for no links, no meta-commentary or explanation of the task, and no reprinting of the prefix. The joined tweet must stay within 280 characters. An invented hashtag is a minor flaw, smaller if it sounds like the author's playful voice, and must never outweigh failures of voice or seam."}]}, "target_words": 19, "reference": "that the emperor wears no clothes, but they bow down to him anyway, its better than being alone ♪", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439816227720937473", "author_group_sha256": "a2a2f63b15a4eed48a85f6079e9641578cafbbf5c7d4c917449318dcd7cc5780", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_2629916247a3cd8095e4e8f8", "id": "2629916247a3cd8095e4e8f8", "domain": "tweets", "language": "en", "split": "train", "group": "btc:476941970859102208", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 23 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nwhen i'm wealthy i'm going to ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (e.g., \"i'm going to [continuation start]\") before deciding. The prefix ends on 'to', so the continuation must begin directly with a verb or verb phrase completing 'going to ...' (e.g., 'buy', 'hire'), with no capitalized restart, no re-typing of 'to' or 'going to', and no sentence-initial hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The completed tweet must read as one coherent plan or fantasy."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, lowercase, daydreaming or joking 'when i'm rich' register. Slang, fragments, run-ons with '..', slashes and exclamations are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives (equivalents of 'perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or moralizing closers such as 'because money can't buy happiness.' A hedged, formal rewrite of an idea scores worse than a rough but lively one."}, {"id": "c3", "description": "Substance: Reward a specific, distinctive wealthy-person fantasy, such as a named thing, a concrete absurd purchase or action, or an actual joke or punchline. Generic stock ideas like 'buy a big house and travel the world' or plain hyperbole count as weak, and cap naturalness around 8 when nothing distinctive accompanies them. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output contains no links, no meta-commentary, no explanation of the task, and no quotation of the prompt. The full tweet including the prefix stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 23, "reference": "have a butler that only works one day a week.. and i'm going to call him reginald even if that's not his name.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "476941970859102208", "author_group_sha256": "f0597b30001746a583f01c652fe1728bdd3c1607313f09a85c6e0ebf96d3a3a4", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_30d6492e5fb41e5f1cfa2118", "id": "30d6492e5fb41e5f1cfa2118", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458976315740549120", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 14 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nGirls got a good way of ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'Girls got a good way of ' (trailing space after 'of'), so the continuation must begin with a gerund or noun phrase that completes 'a good way of ...' (e.g., a verb in -ing form). The judge must quote the joined text at the seam (e.g., 'good way of [first words]') before deciding. Fail if it re-types 'of' or earlier prefix words, starts with a capitalized restart or a sentence-initial hedge, or reads ungrammatically at the join. A continuation that starts right where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, slangy, observational register of 'Girls got a good way of' (dropped 'have', a conversational generalization). Lowercase, slang, fragments, '..' run-ons, nonstandard spelling, or exclamations are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete behavior, image, named thing, or actual joke/punchline about what girls have 'a good way of' doing, such as a recognizable relatable action. Generic phrasing or stock hyperbole (e.g., 'making everything better', 'ruining your day') with nothing distinctive caps naturalness at about 8/10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation text. It has no links, no meta-commentary or explanation of the task, and no quotation marks wrapping the output. The full tweet including the prefix stays within 280 characters. A hashtag in the author's own playful voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 14, "reference": "making a guy feel like he fucked with.. But that don't be the case", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458976315740549120", "author_group_sha256": "54e5cec4e2918d9dc12fb1045a79779726a955278b929c7ceccf96c9fce3da75", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_31d633f0ff550875e3bf2216", "id": "31d633f0ff550875e3bf2216", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463993160600129536", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 14 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nDistance in relationship can cause you ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'can cause you ' (trailing space, expecting an object/verb phrase like 'to...' or a noun). Judge must quote the joined text at the seam (e.g., '...can cause you [continuation start]') before deciding. Pass if the continuation flows grammatically from 'cause you' without re-typing 'you' or 'cause', without a capitalized restart or sentence-initial hedge. A continuation beginning directly with e.g. 'to overthink...' or 'so much stress...' is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: Continuation keeps the casual, personal, slightly emotional/advice-tweet register about relationship distance (lowercase, slang, fragments, '..', emoji-like feeling are fine). Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, or explanatory/moralizing closers like 'so communication is key'. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific image, concrete action, or genuine feeling/joke tied to distance in a relationship (e.g., staring at their last text, overthinking every late reply, missing someone's voice). Stock phrasing like 'so much pain' or 'to lose everything' with nothing distinctive caps naturalness around 8/10; filler or padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): No links, no meta-commentary or task explanation, full tweet within 280 characters. An author-voiced playful hashtag is only a minor flaw and must never outweigh seam or voice failures."}]}, "target_words": 14, "reference": "miss someone so much It helps to realize how much that person means.\n2", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463993160600129536", "author_group_sha256": "3c54564856b1bd676ffce2ab7c46f61ea24d7cba4768ca896101361fb58dd731", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_33c53c5c453068e74cd5bdac", "id": "33c53c5c453068e74cd5bdac", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489769813448093696", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 15 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nFinished watching the first couple of ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'first couple of ' (trailing space, after the preposition 'of'), so it needs a noun phrase next, such as 'episodes of X' or 'eps'. Quote the joined text at the seam, e.g. '...first couple of [continuation start]', before deciding. The continuation must begin directly with that noun phrase. It must not re-type 'of' or 'couple', must not restart with a capitalized sentence, and must not open with a hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The rest must read as one coherent tweet about having just watched something."}, {"id": "c2", "description": "Voice and stance: The continuation keeps a casual, first-person viewer reaction, like a real person tweeting after a watch session. Informality, lowercase, slang, fragments, '..' run-ons, exclamations or emoji-like enthusiasm are fine when they fit. Penalize heavily any assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, review-like summaries, or a tidy moral or explanatory closer. A rough, lively reaction should score above a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward concrete specifics such as a named show, a character, a specific scene, a concrete reaction like 'stayed up till 3am', or an actual joke. Generic praise or stock hyperbole without anything distinctive ('it was amazing, can't wait for more') is weak, and naturalness should be capped around 8/10 in that case. Do not credit filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output contains no URLs, no meta-commentary, and no explanation of the task. The full tweet including the prefix stays within 280 characters. A playful, in-voice hashtag is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 15, "reference": "Billy & Mandy seasons nonetheless (save Jacked-Up Halloween) and I still totally dig this show!", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489769813448093696", "author_group_sha256": "0fe7d1617f5007f759046f7d972b431050c77585e81c54b42b49f552c5dcb4a6", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_39569390f43246ab7a205930", "id": "39569390f43246ab7a205930", "domain": "tweets", "language": "en", "split": "validation", "group": "btc:476939777246515200", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 19 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nHere's to the crazys the misfits ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with \"the misfits \" (trailing space) inside an unfinished list echoing the 'Here's to the crazy ones' toast. Quote the joined text at the seam, e.g. \"...the misfits [continuation]\", before deciding. The continuation should carry the list or toast forward directly, with items like 'the rebels', 'the troublemakers' or a 'who/that' clause. It must not re-type 'the misfits' or 'Here's to'. It must not open with a capitalized restart or a sentence-initial hedge like 'Perhaps' or 'Indeed'. Starting right after 'misfits' is correct and is not repetition."}, {"id": "c2", "description": "Voice: The continuation keeps the casual, celebratory, defiant toast tone of the prefix, including its loose grammar ('crazys'), missing commas and run-on energy. Lowercase, fragments, slang, '..' and exclamations are fine. Penalize heavily any assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, or explanatory or moralizing closers such as 'because diversity makes the world better.' A rough but lively line beats a smooth, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named group, concrete action or real twist or joke. Examples include naming the author's own crew or a particular weird habit, or giving the toast an unexpected personal punchline. A string of stock Apple-ad phrases or generic hyperbole ('the ones who change the world', 'stay weird') with nothing distinctive should score no higher than about 8/10 for naturalness. Do not credit filler clauses that only pad the list."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer contains no links, no meta-commentary, and no explanation of the task or quotation marks framing the output. The full tweet stays within 280 characters. A hashtag is a minor flaw if it sounds like the author's playful voice. It must never outweigh seam or voice failures."}]}, "target_words": 19, "reference": "the rebels the trouble makers the one who see things differently and have no respect for the status quo.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "476939777246515200", "author_group_sha256": "0ed8c64448dfe3cbcf9c79d14579a0e77c46dde3dbc6f3d3bcd46327a562dcfb", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_3b0fc60a9272d3644413dbac", "id": "3b0fc60a9272d3644413dbac", "domain": "tweets", "language": "en", "split": "validation", "group": "btc:489769423356444672", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 14 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nFacebook had a purge last night, ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (\"...purge last night, \" + continuation) before judging. The prefix ends after a comma and space mid-sentence, so the continuation should continue the clause directly, typically lowercase, without re-typing 'purge' or 'last night', without a capitalized restart, and without opening with a hedge like 'Apparently' or 'Perhaps'. A continuation that picks up exactly where the prefix stops is correct and must not be marked as repeating. Penalize joins that are ungrammatical or that make no sense with 'Facebook had a purge last night'."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, wry, social-media-user register of someone reacting to a Facebook purge, whether friends unfriended, accounts banned, or a feed cleanup. Informality, lowercase, slang, '..' run-ons, fragments and exclamations are fine. Heavily penalize assistant-style writing: bookish connectives or hedges ('certainly', 'which is telling', 'it seems'), balanced polished clauses, tidy morals, or explanatory or moralizing closers about social media. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete detail, such as a specific consequence (e.g., a friend count dropping, a named type of person or page vanishing, an aunt getting banned), a concrete action, or an actual joke or punchline. Generic phrasing or stock hyperbole (e.g., 'chaos everywhere', 'crazy times') is weak, and naturalness should be capped around 8/10 when nothing distinctive accompanies it. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that there are no links, no meta-commentary or task explanation, and no quotation-wrapped output, and that the full tweet is at most 280 characters. Hashtags violate the instruction, but one in the author's playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 14, "reference": "there were nudes all over my home page! 😳 😜 😂 😂 👏 👌", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489769423356444672", "author_group_sha256": "de342465345c9adbd31d1faa6d1f3e0dfb902f0016e08620cf1f66866aff623f", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_4001647a3a8073b36a59ac04", "id": "4001647a3a8073b36a59ac04", "domain": "tweets", "language": "en", "split": "test", "group": "btc:476941803091152896", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 19 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nThis girl getting kissed by her ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'kissed by her ' (a possessive needing a noun). Quote the joined text at the seam (e.g., '...kissed by her [first words]') before judging. Pass if the continuation begins with a noun or noun phrase completing 'her' (e.g., a person, pet or thing) and the resulting sentence reads coherently. Fail if it re-types 'her' or earlier prefix words, opens with a capitalized restart or sentence-initial hedge, or leaves the possessive grammatically dangling. A continuation starting directly with the noun is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, reactive tweet register of someone commenting on a scene (amused, awed, envious or teasing). Lowercase, slang, fragments, '..' run-ons, exclamations and enthusiasm are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory/moralizing closers. A rough but lively reaction outranks a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action or actual joke (e.g., a particular kisser like a dog or grandpa, a concrete reaction, a punchline). Stock hyperbole or generic phrasing ('is so cute I can't', 'goals fr') with nothing distinctive caps quality around 8/10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): No links, invented URLs, meta-commentary or task explanation; the full tweet including the prefix is at most 280 characters. A hashtag that fits the author's playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 19, "reference": "mans, but when he stops, she keep making strong eye contact with me. This too funny bruh 😂 😂", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "476941803091152896", "author_group_sha256": "078c9e889012cc902e4d7b886189c7e88860a16f8187cf3b7969f1bfaa1510ef", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_41410ac20eabfc76e376df4c", "id": "41410ac20eabfc76e376df4c", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463986613152784385", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nWind 2,5 km/h N. Barometer 1015,9 ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam, i.e. 'Barometer 1015,9 ' plus the opening of the continuation. The prefix ends after a number and a trailing space, so the continuation should naturally supply a unit (e.g. hPa/mb), a trend word, or the next weather field. It must not re-type '1015,9' or 'Barometer', and it must not open with a capitalized restart or hedge that breaks the reading-list flow. A continuation that picks up exactly at this point counts as correct and is not repetition."}, {"id": "c2", "description": "Voice and stance: The prefix is a terse automated or hobbyist weather-station report with comma decimals and abbreviated units. Reward a continuation that keeps that clipped, data-log register, such as more readings (temp, humidity, rain, trend like 'Rising slowly') or a brief casual aside that fits. Heavily penalize assistant-style prose, including hedges ('perhaps', 'apparently'), bookish connectives, polished balanced clauses, or a summarizing or explanatory closer ('which suggests calm conditions')."}, {"id": "c3", "description": "Substance: Reward specific, plausible concrete details that are consistent with the established format, such as a unit, numeric readings with comma decimals, a named condition or trend. Generic filler ('nice weather today', 'stay safe everyone') or stock phrasing without any distinctive detail caps the quality around 8/10 or lower. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (minor weight): The answer has no links, no meta-commentary or explanation of the task, and the full tweet stays within 280 characters. A hashtag that fits the station-feed voice is only a minor flaw and should never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "hPa, Rising slowly. Temperature 13,3 °C. Rain today 0,0 mm. Humidity 10%", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463986613152784385", "author_group_sha256": "bbac19036de6cecbc794a3c8d501ce23cb20599e200ca10faaa2f9050a027410", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_43546e61e7f6ee874f7907a9", "id": "43546e61e7f6ee874f7907a9", "domain": "tweets", "language": "en", "split": "train", "group": "btc:489770383860846592", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 16 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nTiny perfect storm for #CAD this ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'this ' (a determiner followed by a trailing space), so it needs a noun or time phrase such as 'morning', 'week' or 'Monday'. The judge must quote the joined text at the seam (e.g. '...for #CAD this [continuation start]') before deciding. Pass if the continuation starts directly with a fitting word that completes 'this', with no repeated 'this', no capitalized restart, and no sentence-initial hedge. A continuation that begins exactly where the prefix stops must not be marked as repeating. Fail if the joined text is ungrammatical or the continuation restarts the sentence."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, market-chatter tone of a trader tweeting about #CAD (the Canadian dollar). Fragments, slashes, '..' run-ons, lowercase, slang and exclamations are acceptable. Penalize heavily for hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: The continuation explains the 'perfect storm' with concrete, specific drivers, such as named data releases (jobs report, CPI), oil price moves, BoC/Fed decisions, USD strength, or a level or pair like USDCAD. Generic hyperbole or vague phrasing ('things are crazy', 'lots going on') without a distinctive detail caps the naturalness score at about 8. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains no links, no meta-commentary or task explanation, and the full tweet stays within 280 characters. Treat any new invented hashtag as a minor flaw that never outweighs a voice or seam failure."}]}, "target_words": 16, "reference": "morning, with strong capital inflows joining weak US starts data to bolster the wobbly currency #forex", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489770383860846592", "author_group_sha256": "becc2683d288d18a19e4e829ae2e773d9e42b7756d2415035a30a34b3f471832", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_476519aaf20af6aa60abfb53", "id": "476519aaf20af6aa60abfb53", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439808296003063810", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 24 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nWe do have a favourable run ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'We do have a favourable run ' followed by a trailing space, so the noun phrase is open. Before deciding, the judge must quote the joined text at the seam, e.g. 'favourable run [continuation]'. The continuation should extend the phrase naturally, for example with 'of fixtures', 'coming up' or 'in' plus a time or opponents, and form one coherent standalone tweet. Penalize these failures: re-typing 'run' or other prefix words, starting with a capitalized restart or a sentence-initial hedge, or producing text that doesn't follow grammatically. A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation should keep the casual fan register and the cautiously hopeful or optimistic stance implied by 'We do have'. It may be informal, use lowercase, slang, fragments, '..' run-ons or exclamations. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively fan voice beats a formal, hedged rewrite."}, {"id": "c3", "description": "Substance: Reward concrete specifics such as named or implied opponents, a number of games, a realistic fan worry or boast ('if we don't bottle it'), or an actual joke. Generic stock phrasing ('let's see what happens', 'exciting times ahead') with nothing distinctive is weak and caps the score around 8 of 10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene, a minor weight below c1–c3: There should be no links, no meta-commentary or task explanation, and the full tweet must stay at most 280 characters. Invented hashtags are a minor flaw, and only a small deduction if they sound like the author's playful voice. Never let them outweigh a failure of seam or voice."}]}, "target_words": 24, "reference": "in as well. We go on a winning streak and those above draw a couple of games and we're right in the mix. #pnefc", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439808296003063810", "author_group_sha256": "2dce3e364d9d3e04174ea0e34d12caa541cdff52bc6e5169373622e7cb942c99", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_477036cb7fe7ac0540690216", "id": "477036cb7fe7ac0540690216", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489758987911319552", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 15 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nGroggily got out of bed. Blindly ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (e.g., 'Blindly [continuation]') before judging. The prefix ends on the adverb 'Blindly ' followed by a space, so the continuation should begin with a verb or verb phrase that 'Blindly' modifies (e.g., 'stumbled', 'groped for'). It must not re-type 'Blindly', must not open with a capitalized restart or a sentence-initial hedge, and must read as one grammatical sentence across the seam. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the groggy, casual, first-person-diary morning mood, with its terse, fragment-friendly rhythm (subject-dropped sentences like the prefix). Informal touches, lowercase, slang or fragments are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives (perhaps, apparently, certainly, which is telling), tidy summarizing morals, balanced polished clauses, or explanatory or moralizing closers. A rough but lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete half-asleep action, object or mishap (e.g., stubbing a toe on a named thing, pouring orange juice into coffee, hunting for glasses) or an actual small joke or payoff. Treat stock hyperbole or generic phrasing ('started my day', 'it was a struggle') as weak, and cap the score around 8/10 when nothing distinctive accompanies it. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation has no links, no meta-commentary or explanation of the task, and no quotation of the prompt, and the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 15, "reference": "tried to push window closed. Window met resistance & wouldn't close. Cat sitting in there.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489758987911319552", "author_group_sha256": "7d6578951e1d07ecbe6a5f0ed2f27e47b27969cf0b8d20cde97d38e2e3b96e74", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_54e5637e766e05185a08cdae", "id": "54e5637e766e05185a08cdae", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463995261707358208", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 19 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI am Alice I'm in Wonderland ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (\"...I'm in Wonderland\" + start of continuation) before judging. The prefix ends after a complete clause with a trailing space, so the continuation may begin a new thought (lowercase or capitalized both acceptable in this casual voice) but must not re-type 'Wonderland' or any prefix words, and must read as a natural follow-on to the Alice/Wonderland self-declaration rather than an unrelated restart. A continuation that simply continues from where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the playful, casual, slightly whimsical first-person voice of someone declaring they're Alice in Wonderland (e.g., excited, dazed, joking about something surreal in their day). Informal fragments, lowercase, run-ons, slang and exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, or explanatory/moralizing closers. A rough but lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action or actual joke (e.g., a reference to a rabbit hole, tea party, Cheshire cat, shrinking/growing, or a concrete real-life situation framed as Wonderland). Stock hyperbole or generic phrasing ('everything is so magical and amazing') is weak; cap naturalness around 8/10 when nothing distinctive accompanies it. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains no links, no meta-commentary or explanation of the task, no reprinted prefix, and the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 19, "reference": "Where's the rabbit? He is late again Goodness\nGracious I can't wait for him Who has got the time?", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463995261707358208", "author_group_sha256": "dec5c8c1d3f564ce88be79c8da658e57783ec5d6e5c8e862995037072affbfc5", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_57a9830d33519b307c8c7810", "id": "57a9830d33519b307c8c7810", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458977045448761345", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI hate when people ask me ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'ask me ' (trailing space after 'me'), so the continuation should begin with the object of the asking (e.g., a question, 'what', 'if', 'to', 'about', or a quoted question). The judge must quote the joined text at the seam (e.g., 'I hate when people ask me [continuation start]') before deciding. Penalize continuations that re-type 'me' or 'ask me', open with a capitalized sentence restart or hedge, or produce an ungrammatical join. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the annoyed, casual complaint register of 'I hate when people ask me'. Lowercase, slang, fragments, '..' run-ons, nonstandard spelling, or exasperated exclamations are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory or moralizing closers such as 'it's just not okay' or 'people should respect boundaries'. Rank a rough but lively gripe above a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, recognizable annoying question or situation, or an actual punchline, such as a concrete question quoted in the author's voice or a sharp, specific reason it's irritating. Treat generic complaints ('stupid questions', 'personal stuff', 'things I don't know') or stock hyperbole without a distinctive detail as weak and cap them around 8 for naturalness. Do not credit filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that the output contains no links, no meta-commentary or explanation of the task, and that the full tweet fits within 280 characters. A hashtag that sounds like the author's own playful voice is a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "what happened to my arms or what's on my arm 👿 👿", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458977045448761345", "author_group_sha256": "9bac10cc0d0c4b97db390d37b85f58025b6170ee5aa7e5b8c0f384473ef20198", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_5979b85ad437f338d448ba47", "id": "5979b85ad437f338d448ba47", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458982922004230144", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 24 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nDid you ever love me like ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'like ' (a preposition/conjunction awaiting its object or clause). Quote the joined text at the seam (e.g., 'Did you ever love me like [first words]') before judging. The continuation must supply what is being compared to (a noun phrase or clause such as 'I loved you' or 'the way...') without re-typing 'like', without a capitalized restart or sentence-initial hedge, and must read as one grammatical, coherent tweet. A continuation that begins exactly after 'like ' is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The completed tweet keeps the raw, wistful/accusatory emotional register of a personal question to an ex or lover, in casual tweet voice (lowercase, fragments, slang, '..' run-ons, rhetorical follow-ups are fine). Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, or explanatory/moralizing closers like lessons about self-worth. A rough but heartfelt line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action, or actual wit (e.g., a particular shared memory, habit, or object) that makes the comparison vivid. Stock lines ('like I loved you', 'with all my heart') or generic hyperbole with nothing distinctive should cap naturalness around 8/10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Output contains only the continuation text — no links, no invented hashtags, no meta-commentary or task explanations, no quotes around the answer — and the full tweet with prefix stays within 280 characters. A hashtag in the author's own playful voice is a minor flaw that must never outweigh voice or seam failures."}]}, "target_words": 24, "reference": "I ever love you ? It's my mistake but when I do right nobody remember it , when I'm wrong they remember it frv", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458982922004230144", "author_group_sha256": "f0c303276f85c6e61b2f19acda1a1a8138043cb9611e61703e0c5c4188c15d1e", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_5e92e24e3edc95548568c827", "id": "5e92e24e3edc95548568c827", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463986260961267712", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 16 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nGracious celibate gifts on behalf of ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends on the preposition phrase 'on behalf of ' (trailing space), so the continuation must begin with the object of 'of', such as a name, group, or noun phrase. The judge should quote the joined text at the seam, e.g. '...on behalf of [continuation start]', before deciding. Penalize continuations that re-type 'of' or 'behalf', that open with a capitalized sentence restart or a hedge, or that leave 'of' without an object. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The rest must read as one coherent tweet."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the prefix's wry, oddly formal-yet-casual tweet register. Informal touches such as lowercase, fragments, slang, '..' run-ons, slashes and exclamations are acceptable when they fit. Heavily penalize assistant-style writing, including hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively line ranks above a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific named beneficiary or thing, a concrete image or action, or an actual joke that plays on the 'gracious celibate gifts' idea. Generic phrasing, stock hyperbole, or filler clauses with nothing distinctive count as weak and should cap naturalness at about 8 out of 10. Padding earns no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains no links, no meta-commentary or explanation of the task, and no quotation of the prefix, and the full tweet stays within 280 characters. Hashtags count as a minor flaw even when they sound like the author's own playful voice. This flaw must never outweigh a voice or seam failure."}]}, "target_words": 16, "reference": "absolute people in general against of mark occasions regardless of cost dole out forwarding e: KReOiph", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463986260961267712", "author_group_sha256": "56b3415fc2a4779a831f67c12c084f91b3486a0fa346edbf8f28dfaf270b3498", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_62275670f05ebe1eceb58396", "id": "62275670f05ebe1eceb58396", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489760766296223744", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nY'all don't forget that Manu Ginobili ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with \"Manu Ginobili \" (a subject awaiting its predicate, with a trailing space). The judge must quote the joined text at the seam, e.g. \"...Manu Ginobili [continuation start]\", before deciding. Pass if the continuation begins directly with a predicate or clause about Ginobili (e.g. a verb like \"came off the bench...\" or \"was drafted...\") and reads grammatically as one sentence. Fail if it re-types \"Ginobili\" or \"Manu\", restarts with a capitalized new sentence or hedge (\"Perhaps...\", \"It's worth noting...\"), or leaves the subject without a predicate. A continuation that starts exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The \"Y'all don't forget\" opener sets a casual, hype or reminder-style fan voice, so the continuation should keep that informal, emphatic register. Slang, lowercase, fragments, exclamations, '..' run-ons and enthusiasm are fine. Heavily penalize assistant-style writing: hedges or bookish connectives (\"perhaps\", \"apparently\", \"certainly\", \"which is telling\"), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers (e.g. \"a true testament to dedication\"). A rough but lively line beats a formal, hedged rewrite."}, {"id": "c3", "description": "Substance: Reward a specific, concrete claim or joke about Ginobili, such as the 57th pick, the Euro step, the bat he swatted, sixth man play, Spurs titles, or Argentina's 2004 Olympic gold. Generic praise like \"was a legend\" or \"was the GOAT\" with nothing distinctive counts as weak and caps the score around 8/10. Do not credit filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that the answer has no links, no meta-commentary or explanation of the task, and no quoting of the instructions, and that prefix plus continuation stays within 280 characters. A playful hashtag that sounds like the author's own voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "is the Father of the Flop. He brought \"Flopping\" to the Association.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489760766296223744", "author_group_sha256": "4673b09c3cb85798db4474335927d0082f94c140c7d5606b94480de6c24c1e93", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_72d55184e3362b2c5604bef3", "id": "72d55184e3362b2c5604bef3", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439819796902064128", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 13 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nDon't do legs and cardio while ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'while ' (trailing space after a conjunction expecting a clause or gerund, e.g. 'hungover', 'fasting', 'running on 3 hours sleep'). The judge must quote the joined text at the seam ('...cardio while [continuation start]') before deciding. Pass if the continuation picks up grammatically right there without re-typing 'while' or 'cardio', without a capitalized restart, and without a sentence-initial hedge. Starting directly where the prefix stops is correct and must not be marked as repetition. Fail if the joined text is ungrammatical or restarts the sentence."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, gym-bro or self-deprecating warning tone of 'Don't do legs and cardio while...'. Lowercase, slang, fragments, '..' run-ons, exclamations, and enthusiastic or rueful asides are legitimate. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory/health-advice closers (e.g. 'it's important to listen to your body'). A rough but lively line beats a hedged or formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete situation or consequence that makes the warning funny or vivid. Examples include a named condition (hungover, fasting, after Thanksgiving), a concrete aftermath (crawling up the stairs, legs shaking in the shower, nearly fainting on the treadmill), or an actual punchline. Stock hyperbole ('I literally died', 'worst idea ever') or generic phrasing with nothing distinctive caps the score at about 8/10. Do not credit filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that there are no links, no meta-commentary, no explanation of the task, and no quotation marks wrapping the output. Also check that the full tweet with the prefix stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a seam or voice failure. An invented promotional or generic hashtag counts as a moderate flaw."}]}, "target_words": 13, "reference": "you decide to dance azonto the whole night, I can't even walk lol", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439819796902064128", "author_group_sha256": "c95eb11105886d2e193deb069a2aafd2cec28f40063d2e33b097f80a221e2e37", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_773a1ed14cecb7c12a64b3d8", "id": "773a1ed14cecb7c12a64b3d8", "domain": "tweets", "language": "en", "split": "validation", "group": "btc:439821231463100416", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 22 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nLast chance to come see les ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'les ' (trailing space), so a noun or title is still needed, most likely 'Les Misérables', 'Les Mis' or a French plural noun phrase. The judge must quote the joined text at the seam, for example 'see les [continuation start]', before deciding. The continuation passes if it completes 'les' into a grammatical phrase without re-typing 'les' and without a capitalized restart or sentence-initial hedge. Starting directly with 'mis', 'Misérables' or another fitting completion is correct and must not be marked as repetition. Fail or heavily penalize a continuation that ignores 'les', begins a new sentence, or leaves the phrase broken."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, promotional, excited voice of an author inviting people to a show or event. Lowercase, slang, fragments, exclamations, '..' run-ons and enthusiastic tone are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory closers. A formal, hedged invitation scores worse than a rough but lively one."}, {"id": "c3", "description": "Substance: Reward concrete details that make the invite feel real, such as a specific night or time ('tonight', 'sat matinee'), a venue, a role the author plays, a named song, or an actual joke. Generic hype ('it's going to be amazing', 'don't miss out') with nothing distinctive caps the score around 8 for naturalness. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check for links, invented URLs, meta-commentary about the task, and whether the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw. It must never outweigh failures of seam or voice."}]}, "target_words": 22, "reference": "mis tonight at 7:30! Tickets are $15 at the door. Everyone come see an amazing show and help us sell out!!🇫🇷💙 💙", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439821231463100416", "author_group_sha256": "8e4a75d28bd7da23d96f84fdc8a0ac80fcb1c00a570405883dd974f60825d5c3", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_7bc9de42c1c68e552b9ef1de", "id": "7bc9de42c1c68e552b9ef1de", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489758522356150272", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 20 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nis not a 'he', nor a ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'nor a ' (trailing space, article needing a noun). Quote the joined text at the seam (e.g., '...nor a [first words]') before deciding. The continuation must begin with a noun or noun phrase that completes the 'not a 'he', nor a ___' parallel (e.g., a quoted pronoun like 'she', 'thing', or a playful noun). It must not re-type 'a', 'nor', or 'he', and must not open with a capitalized restart or sentence-initial hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The rest should follow logically from the 'neither X nor Y' setup, e.g., with a punchline, correction, or assertion about what the subject actually is."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, possibly playful, sassy, or corrective tone of a tweet insisting on what someone or something is not. Lowercase, slang, fragments, '..' run-ons, slashes, exclamations, and enthusiasm are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, or explanatory/moralizing closers such as lectures about respecting identity. A rough but lively completion beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action, or an actual joke or twist, such as saying the subject is 'an it, a gremlin who steals my socks' or naming a specific identity or creature. Stock hyperbole or generic phrasing ('just their own person, ok') is weak, and naturalness is capped around 8 when nothing distinctive accompanies it. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that there are no links, no meta-commentary or explanation of the task, and that the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 20, "reference": "'she'. Skylar is a 'she', biologically, but she does not identify as a 'she'. I would really appreciate it (c)", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489758522356150272", "author_group_sha256": "51aaf832febce640293332d2aba2476b4a6885200562b9515b3148ce5e3ca6df", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_7eea5e242b15bf77e09eba34", "id": "7eea5e242b15bf77e09eba34", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458979453201170432", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nOf course leeper gives us a ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'gives us a ' (an article followed by a space), so the continuation must begin with a noun or noun phrase, such as an adjective+noun, that grammatically completes 'a ...'. The judge must quote the joined text at the seam (e.g., 'gives us a [first words]') before deciding. Penalize a continuation that re-types 'a' or 'gives us', opens with a capitalized restart or sentence-initial hedge, or begins with a word that cannot follow 'a' (e.g., a vowel-initial noun requiring 'an' is a minor flaw). Do not mark a continuation as repeating when it starts exactly where the prefix stops."}, {"id": "c2", "description": "Voice and stance: 'Of course' signals a sarcastic, exasperated sports-fan tone about Leeper (likely a player). The continuation should keep that resigned or mocking casual fan voice. Lowercase, slang, fragments, '..' run-ons and exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete payoff, such as a named play type (fumble, pick, penalty, drop, turnover), a game situation, an actual joke, or a vivid image of what Leeper 'gave us'. Generic phrasing or stock hyperbole ('the worst thing ever', 'a total disaster') with nothing distinctive caps naturalness at about 8/10. Do not credit filler or padding clauses."}, {"id": "c4", "description": "Hygiene, weighted below c1–c3: The output contains only the continuation text, with no links, meta-commentary, quotation wrappers, or explanation of the task, and the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 12, "reference": "quiz i get a 100% then tells us its not graded 😑", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458979453201170432", "author_group_sha256": "27f6da36834b864ad6be34a30b9ff7f37062c58af511bec82e1c5ac939e0f722", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_8753fedfc2eb08ef02c6d3bf", "id": "8753fedfc2eb08ef02c6d3bf", "domain": "tweets", "language": "en", "split": "train", "group": "btc:439809659437068288", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 13 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nThat awkward moment when you're answering ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'answering ' (trailing space after a verb needing an object). Quote the joined text at the seam (e.g., 'you're answering [continuation start]') before deciding. The continuation should begin with a natural object or clause completing 'answering' (e.g., a question, a call, someone), in lowercase flow, without re-typing 'answering', without a capitalized restart or sentence-initial hedge. A continuation starting exactly where the prefix stops is correct and must not be marked as repeating. The completed tweet should deliver a coherent 'awkward moment' scenario."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, self-deprecating 'that awkward moment when' meme register: informal, possibly lowercase, fragments, slang, '..' or emoji-like enthusiasm are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, or explanatory/moralizing closers. A rough but lively completion beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete awkward situation or actual punchline (e.g., a named thing, a concrete action like answering a question that wasn't directed at you, or replying to a voicemail greeting). Generic phrasing or stock hyperbole without a distinctive detail caps naturalness around 8/10; do not credit filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, meta-commentary, or explanation of the task, and the full tweet stays within 280 characters. A hashtag in the author's own playful voice is only a minor flaw and should never outweigh failures of seam or voice."}]}, "target_words": 13, "reference": "a test question and the answer is so ovious, you actually doubt yourself.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439809659437068288", "author_group_sha256": "9751cf2131710558ba98c193fadf3c023dcb8084ecc3b8d87f95c1b88fb6e749", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_8a7fa2b550aab5ebcd2e5530", "id": "8a7fa2b550aab5ebcd2e5530", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439811744022941696", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 17 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nCullen's always been right when he ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'when he ' (a subject pronoun awaiting a verb). Quote the joined text at the seam (e.g., '...right when he [first words]') before judging. The continuation must begin with a verb or verb phrase that grammatically completes 'when he', without re-typing 'he' or 'when he', without a capitalized restart or sentence-initial hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The resulting tweet must read as one coherent thought about Cullen being right."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, opinionated fan/friend stance (affirming Cullen's track record) in an informal register. Lowercase, slang, fragments, '..' run-ons, exclamations or emoji-like enthusiasm are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives (equivalents of 'perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, or explanatory/moralizing closers. A rough but lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete detail. This can be what Cullen says or predicts (a named game, team, trade, person, event or situation), a concrete action, or an actual joke or punchline. Generic phrasing such as 'says something' or 'talks' and stock hyperbole are weak, and naturalness is capped around 8/10 when nothing distinctive accompanies them. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, meta-commentary or explanation of the task, and the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 17, "reference": "gives his advice to me, I hate it -.- there's never not been once he's been wrong", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439811744022941696", "author_group_sha256": "d52e6ad4814ffa818213076eb62cf801d4178fa04e806d7d92217938d27d65f6", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_9346a72d402a423cec1e997f", "id": "9346a72d402a423cec1e997f", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439813627109847040", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 14 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nwtf happened to Adrianna last night ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with a trailing space after the complete clause \"wtf happened to Adrianna last night\". The judge must quote the joined text at the seam (e.g., \"...last night [continuation start]\") before deciding. The continuation should read as a natural follow-on to that question, for example by adding detail, reacting, or extending the gossip. It should not re-type \"last night\" or \"Adrianna\" as a restart. It should not open with a hedge or a formal capitalized restart such as \"Perhaps\" or \"It seems\". A continuation that simply continues from where the prefix stops is correct and must not be marked as repeating. A continuation that contradicts the question or is grammatically disjoint fails."}, {"id": "c2", "description": "Voice and stance: The continuation should keep the casual, shocked or gossipy register set by \"wtf\". Lowercase, slang, fragments, '..' run-ons, nonstandard spelling and exclamations are acceptable when they fit. Heavily penalize assistant-style writing, including hedges or bookish connectives (e.g., 'perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and concerned or moralizing closers (e.g., 'hope she's okay, we should all check on our friends'). A rough but lively reaction scores higher than a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete detail about what happened or how the author reacted. Examples include a named place or drink, a vivid action like her crying in an uber or dancing on the bar, or an actual joke. Generic phrasing like 'she was wild' or 'crazy night' with nothing distinctive counts as weak, and naturalness should be capped around 8. Do not credit filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The continuation contains no links, no meta-commentary or task explanation, and no quoting of the instructions. The full tweet must stay within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 14, "reference": "?? She doesn't even have a phone where someone could contact her wtf man", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439813627109847040", "author_group_sha256": "0497970ef467c78834c14d5b10953361a818789676e5c3255277aabea0a524f1", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_95053bc53763831410e72e5e", "id": "95053bc53763831410e72e5e", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439812175691915265", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 17 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nCincinnati is just an awful offensive ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends on the adjective 'offensive ' (with a trailing space), which needs a noun such as 'team', 'line', 'coordinator' or 'unit'. Quote the joined text at the seam, e.g. 'awful offensive [continuation start]', before deciding. Pass if the continuation begins by supplying that noun or a noun phrase and reads as one grammatical tweet. Fail if it re-types 'offensive' or 'awful', opens with a capitalized restart or a sentence-initial hedge, or leaves the adjective dangling. A continuation that starts exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the frustrated, blunt sports-fan register and the negative stance on Cincinnati's offense. Casual slang, lowercase, fragments, '..' run-ons and exasperated exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. Also penalize softening or reversing the criticism. Rank a rough, lively rant above a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward concrete football detail or an actual joke, such as a specific failure (three-and-outs, dropped passes, bad play-calling, punting on 4th and 1), a named player or coordinator, or a vivid comparison. Generic complaints or stock hyperbole without anything distinctive ('they're so bad it's unreal') are weak; cap naturalness around 8 in that case. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation contains no links, no meta-commentary, and no explanation of the task, and the full tweet fits within 280 characters. A hashtag that sounds like the fan's own playful voice is only a minor flaw. It must never outweigh a failure of voice or seam."}]}, "target_words": 17, "reference": "basketball team. A team gets hot against them in Tournament and they are out real quick. #MarchMadness", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439812175691915265", "author_group_sha256": "323c0f55deb528bfff70e336ccd5066267a31fe97dac9b00db85126aebf0f7af", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_9a9238c6549a8f4eae086f4e", "id": "9a9238c6549a8f4eae086f4e", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489764281152733185", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\n#FunFact #Independent is self-supporting: not forced ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'not forced ' (trailing space after 'forced'). Quote the joined text at the seam (e.g., '...not forced [continuation start]') before judging. The continuation should pick up grammatically right there, typically with a completion like 'to rely on...', 'by anyone...', or 'into...'. It must not re-type 'forced' or 'not forced', and it must not restart with a capitalized new sentence or a hedge. A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the playful, quasi-definitional '#FunFact' tone and its pro-independence stance in a casual tweet register. Fragments, exclamations, lowercase and slang are fine. Penalize heavily for assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which shows'), balanced polished clauses, tidy summarizing morals, or explanatory or moralizing closers. A rough but lively finish beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action, or an actual joke or twist on the definition, such as a pointed jab, a punchline, or a concrete example of being 'forced' or dependent. Generic stock phrasing or empty hyperbole like 'to depend on anyone ever!!' with nothing distinctive should be capped at about 8/10 for naturalness. Give no credit for filler or padding clauses."}, {"id": "c4", "description": "Hygiene (minor weight): Check that the output has no links, no newly invented hashtags, no meta-commentary or explanation of the task, and that the full tweet, prefix included, stays at or under 280 characters. A hashtag that fits the author's playful voice is a minor flaw only. It must never outweigh failures of seam or voice."}]}, "target_words": 12, "reference": "to rely on another for money or support-to function/operate alone #WeDV8 #1000Network", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489764281152733185", "author_group_sha256": "9eb7c1ab850ee9bda367a63fcca71ae47d8478fe14dcf1dbbc28fc8319ffdfd5", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_9a9f97d12a82855d7ca0e5f8", "id": "9a9f97d12a82855d7ca0e5f8", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463985208287821824", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nNot what one did, but how ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'but how ' (trailing space, on the interrogative/adverb 'how'). Quote the joined text at the seam (e.g., '...but how [first words]') before deciding. The continuation must pick up directly after 'how' and complete the clause grammatically (e.g., a subject/verb like 'one did it' or an adjective/adverb phrase). Penalize re-typing 'how' or other prefix words, a capitalized restart, or a sentence-initial hedge. A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the aphoristic-but-casual tweet register and the stance that manner matters more than the deed. Informal touches (lowercase, fragments, '..', slang, exclamations) are fine if they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, or explanatory or moralizing closers. A rough but lively completion beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action, or actual joke or twist that gives the 'how' some bite, such as a concrete example of manner or an ironic personal turn. Treat stock phrasing ('...how they made you feel', generic platitudes) or padding as weak, and cap naturalness around 8/10 when nothing distinctive accompanies it. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The answer contains no links, no meta-commentary or explanation of the task, and the full tweet stays at or under 280 characters. A hashtag that sounds like the author's own playful voice is a minor flaw only. It must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "one made feel us. That's the #Leadership. #leadershipdevelopment #compassion #leader #relationship #inspirational", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463985208287821824", "author_group_sha256": "e972e5af630f537c767d605691c8417c7d2c4d9bdb5006616783123b341b8452", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_a1020d83daa0ea5f9fefa025", "id": "a1020d83daa0ea5f9fefa025", "domain": "tweets", "language": "en", "split": "train", "group": "btc:463989448552640512", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 13 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nWind 0 km/h S. Barometer 1023.3 ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text where the prefix 'Barometer 1023.3 ' meets the continuation before judging. The prefix ends after a number and a trailing space, so the continuation should plausibly carry on from that point. Good starts include a unit or trend such as 'hPa, rising' or 'hPa steady', or a new reading. A capitalized restart, a hedge opener, or re-typing '1023.3' is a seam failure. A continuation that picks up exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation should keep the terse, automated weather-station style of the prefix. That means clipped readings, units and abbreviations, such as 'hPa, Rising. Temperature 12.4 °C. Rain today 0.0 mm. Humidity 78%'. A brief casual fragment is acceptable only if it fits that terse style. Penalize heavily anything that reads like an assistant: hedges such as 'perhaps' or 'apparently', polished balanced sentences, explanatory remarks about what the readings mean, or a tidy moral or summary closer."}, {"id": "c3", "description": "Substance: Reward concrete, specific data points or details, such as plausible units, trend words, temperature, humidity, rain amounts or time stamps. These should be internally consistent with calm wind and high pressure, for example no storm claims. Treat generic filler like 'nice weather today' or stock hyperbole as weak, and cap the score around 8 when nothing specific appears. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer must contain no links, no meta-commentary, and no explanation of the task. The combined tweet must stay within 280 characters. Hashtags are a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 13, "reference": "hPa, Rising slowly. Temperature 15.5 ºC . Rain today 0.0 mm. Humidity 99%", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463989448552640512", "author_group_sha256": "3bd997ace501c28e5160e16491090fe2104e290368b7ac265dff264e2466032c", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_a822455d7a84c2ae5d84e938", "id": "a822455d7a84c2ae5d84e938", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439821424543674368", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 18 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI notice lately municipal politicians seem ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'seem ' (a verb expecting a complement such as an adjective, 'to'-infinitive or 'like' phrase). Before deciding, the judge must quote the joined text at the seam (e.g., '...politicians seem [first words]'). Pass if the continuation grammatically completes 'seem' (e.g., 'way more...', 'to think...', 'obsessed with...') without re-typing 'seem' or other prefix words, and without a capitalized restart or sentence-initial hedge. A continuation that simply starts where the prefix stops must not be marked as repeating. Fail if the joined sentence is ungrammatical or incoherent at the seam."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, observational, mildly skeptical or wry tone of someone noticing a trend in local politicians. Informal touches like lowercase, slang, fragments, '..' run-ons or exclamations are fine. Heavily penalize assistant-style writing, including hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers about democracy or civic duty. A rough but lively take beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete observation about municipal politicians, such as a named behavior or thing (potholes, bike lanes, ribbon cuttings, Facebook posts, zoning fights, parking), a concrete action, or an actual joke or punchline. Generic phrasing ('more out of touch than ever', 'to not care about people') or stock hyperbole with nothing distinctive caps naturalness around 8/10. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation text, with no links, meta-commentary, quotation marks wrapping the answer, or explanation of the task, and the full tweet stays within 280 characters. An invented hashtag is a minor flaw if it sounds like the author's playful voice and should never outweigh a voice or seam failure."}]}, "target_words": 18, "reference": "to be using Twitter more. Will be fun to watch them stop using it after the election. #tbay", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439821424543674368", "author_group_sha256": "208f01376ccf0561e0076ad98f1428cb5107ca5adb4bc86c6ea14ac4726ecc22", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_aea545049181b3637132097b", "id": "aea545049181b3637132097b", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458969869221126144", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 15 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\n#Fitness #Instructors in #Milwaukee. Preview our ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with \"Preview our \" (trailing space after a possessive that needs a noun). The judge must quote the joined text at the seam, e.g. \"Preview our [continuation start]\". The continuation should begin directly with the noun phrase being previewed, such as a class, schedule or trainer lineup. It must not repeat \"our\" or \"Preview\" and must not open with a capitalized restart or hedge like \"Perhaps\" or \"Check out\". A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The result must read as one coherent promotional sentence."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the upbeat, punchy, local-business promo register of a hashtag-heavy listing tweet. Short fragments, exclamations and casual enthusiasm are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or moralizing closers about fitness or health. A rough but lively promo beats a formal, hedged rewrite."}, {"id": "c3", "description": "Substance: Reward concrete specifics such as a named offering (e.g. 'spin & HIIT instructors', 'trainer profiles'), a concrete action (book, browse, try a free class), or a Milwaukee-specific touch. Stock phrasing alone ('top-notch fitness experts', 'amazing results') is weak and caps naturalness around 8 unless something distinctive accompanies it. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (minor weight): The output contains no invented URLs or links, no meta-commentary or task explanation, and no quotation of the prefix. The full tweet stays within 280 characters. Added hashtags break the instruction and are a small flaw, but must never outweigh a failure of voice or seam."}]}, "target_words": 15, "reference": "summer collection on Thursday from 5:30 - 8:30pm. Instructors get 30% off full priced items", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458969869221126144", "author_group_sha256": "3e827d5cc1968b94996c2220e6af5b9c5ae1d51305f1c8628fd30170f2d207c1", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_bc795e33ba0dd71a2e148a1f", "id": "bc795e33ba0dd71a2e148a1f", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439819582883131393", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 15 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nLydia's been complaining, can't imagine when ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: Quote the joined text at the seam (\"...can't imagine when\" + continuation) before deciding. The prefix ends on 'when ' expecting a clause (e.g., 'she'll stop', 'it'll end'). The continuation must start the clause directly, without re-typing 'when' or other prefix words, and without a capitalized restart or a sentence-initial hedge. If it begins exactly where the prefix stops, do not mark it as repeating. Penalize joins that are ungrammatical or don't make sense."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, wry, mildly exasperated tone toward Lydia's complaining. Lowercase, slang, fragments, '..' run-ons and sarcastic exaggeration are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or moralizing closers. Rate a hedged, formal version below a rough but lively one."}, {"id": "c3", "description": "Substance: Reward a specific detail about what Lydia complains about, a concrete action, a named thing, or an actual sarcastic joke or punchline. Stock hyperbole or generic phrasing alone ('it never ends lol', 'so annoying') is weak. Without something distinctive, cap the score around 8/10. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer contains no links, no meta-commentary, and no explanation of the task, and the full tweet fits within 280 characters. A hashtag that sounds like the author's playful voice is only a minor flaw. It must never outweigh a failure of seam or voice."}]}, "target_words": 15, "reference": "she officially gets her braces. She definitely won't handle it 😂 😂 😂 😂 😂", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439819582883131393", "author_group_sha256": "97a6993d54be51787704301df6b481f601d7054d26dfc29bad58ccb96768e4e6", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_bfcebdc1fa389e33017b1464", "id": "bfcebdc1fa389e33017b1464", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489767703700189184", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 16 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nMy dogs such a meat head... ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The judge must quote the joined text at the seam ('...meat head... ' + continuation) before deciding. The prefix ends after a complete clause plus a trailing ellipsis and space. The continuation should read as a natural next thought about the dog: an example, an anecdote or a punchline. It must not re-type 'meat head' or restate the prefix as if starting over. A lowercase start or a casual capitalized new sentence are both acceptable after the '...'. A continuation that simply picks up after the ellipsis is correct and must not be marked as repeating. Penalize continuations that change subject away from the dog or that contradict the affectionate-exasperated stance."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, affectionately teasing owner voice. Informal grammar, missing apostrophes, lowercase, '..' run-ons, fragments and exclamations are acceptable. Heavily penalize assistant-style writing, including hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), polished balanced clauses, and tidy morals or explanatory closers (e.g., 'but I love him anyway, dogs teach us patience'). A rough but lively line should score above a correct but formal one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific dog behavior or image that shows why the dog is a meat head. Examples of the kind of detail that counts: running into the glass door, barking at his own reflection, chasing a leaf for an hour, forgetting where he dropped the ball. An actual joke also counts. Generic phrasing like 'he's so dumb lol' or stock hyperbole with nothing distinctive caps naturalness around 8/10. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains no links, no meta-commentary or explanation of the task, and no quotation of instructions, and the full tweet stays within 280 characters. An invented hashtag is a minor flaw, smaller still if it sounds like the author's own playful voice. It must never outweigh a failure of voice or seam."}]}, "target_words": 16, "reference": "She broke her collar by running straight off her dog run.. Cool.. I own the hulk", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489767703700189184", "author_group_sha256": "045e889c918a153424dfe924216868bad8e89b58bd80aa9ceacdb10f14170141", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_c4df2bcf388f83813a0da5aa", "id": "c4df2bcf388f83813a0da5aa", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489766587990573056", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 22 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nOne time I had this mini ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends on the adjective 'mini ' (with trailing space), which needs a noun. The judge must quote the joined text at the seam (e.g., 'One time I had this mini [continuation start]') before deciding. Pass if the continuation begins directly with a noun or noun phrase completing 'mini' (e.g., 'fridge', 'cactus', 'panic attack'). Pass also if it begins with a compound like 'golden doodle', and the result must read as one grammatical tweet. Fail if it re-types 'mini' or 'this mini', or opens with a capitalized restart or a sentence-initial hedge. Fail also if it leaves 'mini' without a noun. Do not mark a continuation as repeating when it simply starts where the prefix stops."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, anecdotal storytelling register of 'One time I had this mini...'. That means a personal, offhand recollection. Lowercase, slang, fragments, run-ons with '..', and exclamations are acceptable when they fit. Heavily penalize assistant-style writing, including hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'). Also heavily penalize polished, balanced clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively anecdote beats a hedged, formal one."}, {"id": "c3", "description": "Substance: The anecdote delivers something specific: a named object, a concrete action or event involving the 'mini' thing, or an actual punchline or twist. Stock hyperbole ('it was the best thing ever', 'changed my life') or generic phrasing with nothing distinctive caps the score around 8/10 for naturalness. Filler or padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation text. It must have no links, no meta-commentary, no quotation of the instructions, and no explanation of the task. The full tweet including the prefix stays within 280 characters. A hashtag in the author's playful voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 22, "reference": "catfish named Stanley and he was chill af, just did his thing and kept the gravel in my aquarium from getting moldy", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489766587990573056", "author_group_sha256": "c8acc3f683a8e0fcc336556cc8f18669fe65abef2f8a6ac63d3dae77a9a0df40", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_c5e37ef82b78bbedab05a6cd", "id": "c5e37ef82b78bbedab05a6cd", "domain": "tweets", "language": "en", "split": "test", "group": "btc:458982334734147584", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nCome to think of it, it's ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with \"it's \" (trailing space), so the continuation must supply a predicate or complement of \"it's\" (e.g., an adjective, noun phrase or clause). The judge must quote the joined text at the seam (\"...it's [first words]\") before deciding. Penalize re-typing \"it's\", a capitalized restart, or a sentence-initial hedge that breaks the grammar. A continuation that begins directly with the complement is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation should keep the casual, musing 'come to think of it' register of a realization being shared offhand. Lowercase, slang, fragments, '..' run-ons, exclamations and enthusiasm are fine where they fit. Heavily penalize assistant-style writing: hedges or bookish connectives (perhaps, apparently, certainly, which is telling), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific observation, named thing, concrete image or action, or an actual joke or punchline that pays off the 'come to think of it' realization. Generic or stock phrasing such as 'kind of weird' or 'been forever', or stock hyperbole with nothing distinctive, counts as weak and should be capped around 8/10 for naturalness. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that the output contains no links, no meta-commentary or explanation of the task, and no invented hashtags, and that the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw. It must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "bearday tomorrow and i have a date with the wrong bestfriend :))", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458982334734147584", "author_group_sha256": "42271a471117c05c1745b7cfc8ff15a83ff8c04ddd1dbaafca5d3a70f1d457f4", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_c6d82b60b42f4990d8c786a8", "id": "c6d82b60b42f4990d8c786a8", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489771361141661696", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 15 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI'm scared a college is gonna ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'gonna ' (trailing space), so a verb phrase is needed next. The judge must quote the joined text at the seam, e.g. 'I'm scared a college is gonna [continuation]', before deciding. Pass if the continuation starts directly with a fitting verb or adverb (e.g. 'reject me', 'see my grades') and reads as one grammatical sentence. Fail if it re-types 'gonna' or 'is', opens with a capitalized restart or a hedge ('Perhaps...', 'Well,'), or does not follow grammatically. A continuation that begins exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the anxious, casual, teen-applicant register. Lowercase, slang, fragments, '..' run-ons, exaggerated panic and self-deprecation are acceptable. Heavily penalize assistant-style writing: hedges or bookish connectives ('apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and reassuring or explanatory closers such as 'but I know it'll work out' or 'it's important to stay positive'. A rough but lively fear is better than a formal, measured one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete fear or joke tied to the college process, such as a named detail (GPA, essay, the 'we regret to inform you' email, a typo in the application, the dining hall) or a vivid imagined scenario. Treat generic phrasing ('reject me and ruin my life', 'not want me') with nothing distinctive as weak, and cap naturalness around 8. Give no credit for filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, no meta-commentary, no explanation of the task, and the full tweet stays within 280 characters. A hashtag the author would plausibly use playfully is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 15, "reference": "reject me cause I didn't share that photo of Jesus on the cross on facebook.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489771361141661696", "author_group_sha256": "993a39d07d53bf87e83c2d5c03af5baa72741a279c1cca5223b5f503a1981bd4", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_ca26407e0c0d532c371c4ee8", "id": "ca26407e0c0d532c371c4ee8", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489757238899134465", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 18 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nLive daringly, boldly, fearlessly. Taste the ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends on 'Taste the ' (an article needing a noun). Quote the joined text at the seam (e.g., 'Taste the ___') before judging. The continuation must begin with a noun or noun phrase completing 'Taste the', must not re-type 'Taste the' or 'the', and must not open with a capitalized restart or sentence-initial hedge. A continuation starting exactly where the prefix stops is correct and must not be marked as repeating. The result must read as one coherent tweet."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the prefix's punchy, motivational, exhortative register (imperative, triadic rhythm, enthusiastic). Casual touches such as fragments, exclamations, slang or lowercase are fine if they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'certainly', 'which is telling'), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively line beats a hedged or formal one."}, {"id": "c3", "description": "Substance: Reward a specific, vivid image or concrete thing to 'taste' (a named food, place, sensation or action) or an actual joke or twist. Treat stock phrases such as 'taste the freedom/adventure/life' and generic hyperbole as weak, capping naturalness around 8 when nothing distinctive accompanies them. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output contains no links, no meta-commentary or explanation of the task, and no quoted repetition of the prefix, and the full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 18, "reference": "relish to be found in competition in having put forth the best within you. - Henry J. Kaiser", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489757238899134465", "author_group_sha256": "235bd490a39da6a27c578d60ccf3cd87ccde1753b050aa7addd2c81f02e275f7", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_cb095b8805e53e6f677102ce", "id": "cb095b8805e53e6f677102ce", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463982863365308417", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 21 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI try to look away but ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'but ' (a conjunction followed by a space), so the continuation should begin with a clause completing 'I try to look away but ...', such as a subject, verb or 'i can't'. The judge must quote the joined text at the seam, e.g. 'I try to look away but [first words]', before deciding. Penalize continuations that re-type 'but', start with a capitalized restart or a sentence-initial hedge, or fail to form a grammatical, sensible completion of the look-away contrast. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, confessional, slightly helpless tone of someone who can't stop watching or staring at something. Slang, lowercase, fragments, '..' run-ons, exclamations and nonstandard spelling are acceptable when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, and explanatory or moralizing closers. A rough but lively completion scores above a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific object of attention or a concrete image, action or joke that explains why the author can't look away. Examples include a named show, a person, a train-wreck situation, or a vivid physical detail. Treat stock hyperbole and generic phrasing like 'it's just so captivating' as weak, and cap naturalness around 8/10 when nothing distinctive is present. Give no credit for filler clauses or padding."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check for links, invented hashtags, meta-commentary or any discussion of the task, and confirm the full tweet including the prefix is at most 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 21, "reference": "I can't resist\nEvery time I try to call it quits\nHear come some tits, uh huh, thats a big 10-4😁", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463982863365308417", "author_group_sha256": "b2cb61fed5af9076c825b779458275acaf62636033f6b304a3bdf0dd81ba0325", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_ccde3e589d7055874cb82dba", "id": "ccde3e589d7055874cb82dba", "domain": "tweets", "language": "en", "split": "train", "group": "btc:458983378965262336", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 13 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nSome twitts fly after all. Ask ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'Ask ' (trailing space after the verb), so the continuation should begin with the object of 'Ask' (e.g., a person, group, or pronoun) and read as one grammatical sentence. The judge must quote the joined text at the seam (e.g., 'Ask [first words]') before deciding. Penalize re-typing 'Ask', a capitalized restart, or an opening hedge that breaks the sentence. A continuation that starts exactly at the object is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The text keeps the playful, casual register of 'Some twitts fly after all', including its pun on tweets and birds and its nonstandard 'twitts'. Informal fragments, lowercase, slang, '..' or exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy morals, balanced polished clauses, or explanatory closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific addressee, named thing, concrete action, or an actual joke that pays off the 'twitts fly' wordplay, such as someone a tweet reached or a tweet that went viral. Generic phrasing or stock hyperbole with nothing distinctive caps naturalness around 8 out of 10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check for links, meta-commentary, explanations of the task, and whether the full tweet stays within 280 characters. A hashtag in the author's playful voice is only a minor flaw and should never outweigh a failure of seam or voice."}]}, "target_words": 13, "reference": "urself what u have done to advanced urlife since January 1 of 2014...", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "458983378965262336", "author_group_sha256": "45d37dece63da96069de3be2bb55e2f34d1aad510e2fc9cfadde1f3620797a39", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_ce35b90c1ab746188a785ef1", "id": "ce35b90c1ab746188a785ef1", "domain": "tweets", "language": "en", "split": "validation", "group": "btc:489761030537351168", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 23 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nSo, I just had to fire ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'fire ' (trailing space after a verb needing an object). The judge must quote the joined text at the seam (e.g., 'So, I just had to fire [continuation start]') before deciding. The continuation should begin directly with the object of 'fire' (a person, role, pronoun, or thing) or something grammatically valid after it. It must not re-type 'fire', must not open with a capitalized restart or sentence-initial hedge, and must read as one coherent tweet. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, first-person, slightly exasperated or wry tone set by 'So, I just had to...'. Informality, slang, lowercase, fragments, '..' run-ons, exclamations and nonstandard spelling are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy summarizing morals, balanced polished clauses, and explanatory or moralizing closers (e.g., 'a reminder that communication matters'). A rough but lively take beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a specific detail about who was fired or why. Examples include a named role or person, a concrete absurd action they did, a vivid image, or an actual punchline. Generic phrasing like 'someone for not doing their job' or stock hyperbole counts as weak and caps naturalness around 8/10 unless it comes with something distinctive. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output must not contain links, meta-commentary, or any explanation of the task. The full tweet including the prefix must be at most 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw and must never outweigh a failure of voice or seam."}]}, "target_words": 23, "reference": "someone...\nKeep in mind I'm not a supervisor, but was told by a supervisor \"yeah go ahead and take care of that.\" WTF", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489761030537351168", "author_group_sha256": "9353338a34a3f9ffedfe0bdb12ce9ec29356d1ce0ad47bb2bb1b31f1daeec5ac", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_d0b36fcc5c3f5d5e2d63140a", "id": "d0b36fcc5c3f5d5e2d63140a", "domain": "tweets", "language": "en", "split": "test", "group": "btc:489764499251937281", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nThe issue is that the food ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with the noun phrase 'the food ' (trailing space), so the continuation should begin with a predicate or modifier for 'the food', such as a verb ('is', 'was', 'tastes') or a relative clause. Before deciding, the judge must quote the joined text at the seam (e.g., 'the food [first words]'). Penalize a continuation that re-types 'food' or 'the food', starts with a capitalized restart, or opens with a sentence-initial hedge. Do not mark a continuation that starts exactly where the prefix stops as repeating. The completed tweet must read as one coherent complaint about what the issue with the food is."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, complaining stance set up by 'The issue is that'. Informal phrasing, slang, lowercase, fragments, run-ons, and exclamations are fine when they fit. Heavily penalize assistant-style writing: hedges and bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced or polished clauses, tidy summarizing morals, and explanatory or moralizing closers. Rate a rough but lively gripe above a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific problem with the food, such as a named dish, an exact flaw (cold fries, $18 for three bites, tastes like the fridge), a concrete action, or an actual joke. Treat generic phrasing ('is not good', 'is really bad') or stock hyperbole as weak, and cap naturalness around 8 when nothing distinctive accompanies it. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene, weighted below the other criteria: There should be no links, no invented hashtags, no meta-commentary or explanation of the task, and the full tweet must stay within 280 characters. Treat a hashtag that sounds like the author's own playful voice as a minor flaw only. Never let it outweigh a failure of seam or voice."}]}, "target_words": 12, "reference": "at my gym is too good and the workouts are too hard...", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "489764499251937281", "author_group_sha256": "29bdda2cae28e51ce81fd47961decba93a7720fec85ba515d917d10768d55eba", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_d1a46ecfc5bc07d66bdf6716", "id": "d1a46ecfc5bc07d66bdf6716", "domain": "tweets", "language": "en", "split": "test", "group": "btc:439809415999672320", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 19 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nMaking my Oscar picks... a day ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'a day ' (trailing space after 'day'). The judge must quote the joined text at the seam (e.g., 'a day [continuation start]') before deciding. The continuation should grammatically extend 'a day' (e.g., 'late', 'early', 'before...', 'after...'), with no re-typing of 'day' or 'a day', and no capitalized restart or sentence-initial hedge. A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, self-aware Oscar-picks voice of the prefix, which has '...' trailing. Lowercase, fragments, slang, '..' run-ons, exclamations or playful excess are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), tidy morals, balanced polished clauses, or explanatory closers. A rough but lively line beats a hedged formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete detail such as a named film, actor, category, a prediction, or a real joke about being late or early or about guessing. Generic lines like 'can't wait, it's gonna be amazing' are weak, and naturalness is capped around 8 when nothing distinctive accompanies them. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The continuation has no links, no meta-commentary or task explanation, and keeps the full tweet within 280 characters. An author-voice hashtag is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 19, "reference": "before awards. Don't have insights for a full predictions post like last year, but I'll share my ballot later", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439809415999672320", "author_group_sha256": "19e7eb1def9eba4db3c224aec7ca8307339576d7a571cbc29760028bc7eff720", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_d22b659214c56ed853004ccb", "id": "d22b659214c56ed853004ccb", "domain": "tweets", "language": "en", "split": "train", "group": "btc:463996184584269824", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 14 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nShowers this afternoon! Current temp is ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'Current temp is ' (trailing space), so a temperature value or reading is expected next. Quote the joined text at the seam (e.g., '...Current temp is [first words]') before deciding. The continuation should begin directly with a number/temperature or a fitting descriptor, without re-typing 'is' or 'Current temp', and without a capitalized restart or a hedge like 'Perhaps'. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The rest should follow logically from an afternoon-showers weather update."}, {"id": "c2", "description": "Voice and stance: Keep the upbeat, casual, exclamatory weather-update tone of 'Showers this afternoon!'. Informal fragments, lowercase, slang, '..' run-ons, and enthusiasm are fine. Heavily penalize assistant-style writing. This includes hedges or bookish connectives ('perhaps', 'apparently', 'certainly'), polished balanced clauses, tidy morals, and explanatory closers such as 'so remember to stay safe and prepared'. A rough but lively finish beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward concrete, specific details that a real person posting about the weather would add. Examples include an actual temperature figure with a unit or vibe (e.g., '58 and muggy'), humidity or wind, a named plan such as grabbing an umbrella or cancelling a run, or a small joke. Score generic phrasing ('perfect weather for staying in') or stock hyperbole as weak, and cap it around 8/10 when nothing distinctive accompanies it. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output must contain no links, no meta-commentary, and no explanation of the task. The full tweet must stay within 280 characters. Hashtags break the assignment, but a playful one in the author's own voice is only a minor flaw. It should never outweigh a voice or seam failure."}]}, "target_words": 14, "reference": "10c with 9c lows and 12c highs. Winds are around 14mph to the Southwest.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463996184584269824", "author_group_sha256": "1e31cfcea3d11b145c02b5c191fb59e4e1b9cc501252e9182bdca31c8a55bbab", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_d4bb31c890ffeaef94d94184", "id": "d4bb31c890ffeaef94d94184", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463986734812766208", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 22 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nWen wil our girls return?If our ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'If our ' (trailing space after 'our'), so the continuation should begin with a noun or noun phrase completing 'If our ...' (e.g. a noun like 'government' or 'leaders'), not re-type 'our' or 'If', and not open with a capitalized restart or hedge. The judge must quote the joined text at the seam (e.g. '...?If our [continuation start]') before deciding. A continuation that begins exactly at that point is correct and must not be marked as repeating. The completed conditional should read as a coherent clause tied to the question about 'our girls' returning."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's urgent, frustrated, informal tone. The prefix has nonstandard spelling ('Wen wil') and a missing space after '?', so roughness, run-ons, '..', fragments, and emphatic or exasperated phrasing are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory or moralizing closers. A rough but lively demand beats a formal, hedged rewrite."}, {"id": "c3", "description": "Substance: Reward a concrete accusation, named actor or action, or vivid specific detail that makes the plea or anger feel real, such as who failed and what they should do. Generic laments or stock hyperbole without anything distinctive cap naturalness at around 8/10. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, invented hashtags, meta-commentary, or explanation of the task. The full tweet, including the prefix, must be at most 280 characters. A hashtag in the author's own voice is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 22, "reference": "govt claim nt 2 ve d manpower 2 brng dem back,atleast im certain dey have d money 2buy evryone of dem back", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463986734812766208", "author_group_sha256": "77b3361ff66411a1b0a8ede5f186e25f3f5a087e4f35e51598b17e69cdf4cee4", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_e393cea1242689d55d3366bf", "id": "e393cea1242689d55d3366bf", "domain": "tweets", "language": "en", "split": "test", "group": "btc:463991466004451328", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 12 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nI miss secondary school life . ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with a completed sentence ('I miss secondary school life . '). Quote the joined text at the seam before deciding. The continuation should read as a natural next thought, such as elaborating on what is missed or a nostalgic reaction. It must not re-type 'I miss secondary school life' or restate the same idea in different words. A fresh sentence start, lowercase or capitalized, is acceptable here because the prefix ends a sentence. Penalize continuations that contradict or ignore the nostalgic premise."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the wistful, casual, personal register of a real person tweeting nostalgia. Slang, lowercase, '..' run-ons, fragments and exclamations are fine. Heavily penalize assistant-style writing: hedges and bookish connectives (e.g. 'perhaps', 'truly', 'certainly', 'indeed'), tidy morals (e.g. 'cherish every moment', 'you never know what you have until it's gone'), balanced polished clauses, and explanatory or reflective closers. A rough but lively line beats a smooth, formal one."}, {"id": "c3", "description": "Substance: Reward specific, concrete secondary-school details or a real joke. Examples include break time, a named snack, uniforms, dodging homework, a teacher's catchphrase, or 'no bills just vibes'. Generic phrasing such as 'those were the best days' or 'good times' or stock hyperbole without a distinctive detail is weak, and naturalness is capped around 8 of 10. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check for links, hashtags, meta-commentary, or any explanation of the task. Check that prefix plus continuation stays within 280 characters. A hashtag that fits the author's playful voice is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 12, "reference": "That was where the teachers actually gave a fuck about you .", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "463991466004451328", "author_group_sha256": "989b2d4c51f8b0d51428cf129557ca8d30dd567161f2ae71baf63166b4b32260", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_en_f0a06f384bcddd5e1e2e57e7", "id": "f0a06f384bcddd5e1e2e57e7", "domain": "tweets", "language": "en", "split": "train", "group": "btc:439807885082914816", "prompt": "Finish this standalone English tweet naturally, maintaining its stance and casual voice. Do not repeat the prefix, invent a link or hashtag, or discuss this task. Use roughly 17 words; the full tweet including the prefix must be at most 280 characters. Output only the remaining text.\n\nPrefix:\nListening to us lose at Stoke ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'at Stoke ' (trailing space after a complete place name). Quote the joined text at the seam, e.g. 'Listening to us lose at Stoke [continuation start]', before deciding. Pass if the continuation flows grammatically from 'Stoke' (e.g. continuing the clause with 'on the radio', 'while...', ', ...', or a new casual clause) without re-typing 'Stoke' or other prefix words, and without a sentence-initial hedge or a stiff capitalized restart that breaks the run-on feel. A continuation that simply picks up after 'Stoke' is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: Reads like a fed-up football fan tweeting in real time. Stance stays frustrated, resigned or gallows-humorous about the team losing. Informal touches like lowercase, slang, fragments, '..' run-ons and exasperated exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('perhaps', 'apparently', 'certainly', 'which is telling'), balanced polished clauses, tidy morals, or explanatory or moralizing closers such as 'but that's football' or 'we'll bounce back'. A rough, lively line beats a correct, formal one."}, {"id": "c3", "description": "Substance: Rewards concrete, specific detail tied to the situation. Examples include radio or commentary, a cold wet Stoke night, a named or described player blunder, a set piece conceded, doing chores while listening, or a real joke. Generic lines like 'worst day ever' or 'so painful' with nothing distinctive cap naturalness at around 8/10. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Fails if the continuation contains a URL, meta-commentary or explanation of the task, or quotes the prefix, or if the full tweet exceeds 280 characters. A playful author-voice hashtag is only a minor flaw and must never outweigh a voice or seam failure."}]}, "target_words": 17, "reference": "on my TuneIn app at work was like slowly chewing/swallowing a rotten piece of fruit. Feel sick.", "license": "CC-BY-4.0", "source": {"url": "https://raw.githubusercontent.com/GateNLP/broad_twitter_corpus/0c45ef7b1924a6b8115bceab388781590b1b553a/h.json", "revision": "0c45ef7b1924a6b8115bceab388781590b1b553a", "license": "CC-BY-4.0 (corpus declaration)", "tweet_id": "439807885082914816", "author_group_sha256": "7cfd8e47ef843dae08fd3f7b9c8fc679a9731a1fd1ebe28ff5235539f7433664", "quality": "Real historical tweets, not expert writing; human-better prior is especially weak."}} {"task": "tweets_ru_008a8291da9d787981a2013d", "id": "008a8291da9d787981a2013d", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:785", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 19 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nТак ребята вы не представляете какая ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'какая ' (feminine adjective/intensifier awaiting a noun or adjective+noun, e.g. 'какая жесть', 'какая погода'). Judge must quote the joined text at the seam ('...какая [first words]') before deciding. Pass if the continuation starts directly with a word that agrees grammatically (feminine singular noun/adjective phrase) without retyping 'какая' or earlier prefix words, and without a capitalized restart or hedge opener. Starting exactly where the prefix stops is correct, not repetition."}, {"id": "c2", "description": "Voice: Continuation keeps the excited, chatty, addressing-friends tone ('Так ребята вы не представляете') — informal Russian, possibly lowercase, no commas, run-ons, '..', ')))' or exclamations are fine. Heavily penalize assistant-style writing: hedges or bookish connectives (наверное, по-видимому, безусловно, что показательно), polished balanced clauses, tidy morals or explanatory/moralizing closers. A rough lively version beats a formal one."}, {"id": "c3", "description": "Substance: Reward a concrete payoff for 'вы не представляете какая...' — a specific thing, event, image, named place/person, action, or actual joke explaining what was so remarkable. Stock hyperbole alone ('какая крутая вещь произошла, просто нет слов') is weak and caps naturalness around 8/10; filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight): No links, meta-commentary, task explanations or quotation framing; full tweet within 280 characters. Invented hashtags are a minor flaw if playful and in-voice, never outweighing seam or voice failures."}]}, "target_words": 19, "reference": "дичь сегодня на улице! Кто-то выгуливает кота на поводке, а он прямо как собака! Что за мир? #дичь #что_происходит", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 785, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_04ad650d6c1ad4f9d0fad671", "id": "04ad650d6c1ad4f9d0fad671", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1107", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда наконец-то осознал что я ноулайфер ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"...что я ноулайфер\" + continuation) before deciding. The prefix ends after a complete word and a space, inside an open \"Когда...\" clause. The continuation must complete that clause naturally. Good options are a main clause like \"...ноулайфер и даже не расстроился\", or a comma plus consequence, or an emoji-reaction style payoff. It must not re-type \"ноулайфер\" or the prefix. It must not restart with a capitalized new sentence that leaves the \"Когда\" clause dangling. It must not open with a hedge. Lowercase starting directly where the prefix stops is correct and is not repetition."}, {"id": "c2", "description": "Voice: The continuation keeps the casual, self-ironic Russian internet register of someone admitting they're a 'ноулайфер'. Slang, missing commas, lowercase, fragments, ')))' or similar are fine. Heavily penalize assistant-like phrasing. This includes bookish connectives or hedges (\"пожалуй\", \"по-видимому\", \"безусловно\", \"что показательно\"), balanced polished clauses, and tidy morals or reflective conclusions about self-acceptance or balance in life. A rough, lively line beats a correct, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete payoff, such as a specific reaction, image, or joke. Examples: a named game or anime, hours played, mom asking about going outside, ordering doshirak, or deciding to just keep playing. Generic hyperbole or vague phrasing (\"это было ужасно\", \"жизнь изменилась\") with nothing distinctive caps naturalness around 8/10. Padding or filler clauses get no credit."}, {"id": "c4", "description": "Hygiene (lower weight than seam, voice, and substance): The continuation has no links, no meta-commentary or explanation of the task, and no quoting of instructions, and the full tweet fits within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw. Invented or promotional hashtags are a moderate flaw. Neither should outweigh seam or voice failures."}]}, "target_words": 13, "reference": "а не просто «человек с хобби», но это и есть моя суперсила! #НоулайферСтрасть #ЖизньБезСонца", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1107, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_0e502c9762aed301da06d8e5", "id": "0e502c9762aed301da06d8e5", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:931", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЭй, вы видели как наш новый ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'наш новый ' (masculine adjective awaiting a noun). Quote the joined text at the seam (e.g. 'наш новый ___'). The continuation must begin directly with a masculine noun or noun phrase agreeing with 'новый' (e.g. a person, pet, device, coach, player), without re-typing 'наш'/'новый', without a capitalized restart or opening hedge. The remainder must grammatically complete the 'вы видели как...' clause (what this new thing did). A continuation starting exactly at the noun is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: Keeps the excited, chatty, addressing-friends register set by 'Эй, вы видели'. Lowercase, slang, run-ons, '..', exclamations, missing commas are fine. Heavily penalize assistant-like Russian: hedges and bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, tidy morals or explanatory/moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Rewards a specific, concrete image or action — a named or clearly identifiable new thing and something particular it did (a funny mishap, a specific feat), or an actual joke. Stock hyperbole ('просто бомба', 'это нечто') or generic phrasing with nothing distinctive caps naturalness around 8/10; filler or padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, no meta-commentary or explanation of the task, only the continuation returned, and the full tweet fits within 280 characters. Invented hashtags are a minor flaw if they sound playful and authorial; they must never outweigh voice or seam failures."}]}, "target_words": 13, "reference": "план оказался просто козырным? Сначала думал, что провал, а он взял и выстрелил!", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 931, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_14373298dbd12332969167c3", "id": "14373298dbd12332969167c3", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1292", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 10 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nСпам это как неотозванное приглашение на ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends on the preposition 'на ' and requires a noun phrase in the accusative (e.g. 'на вечеринку', 'на свадьбу'). The judge must quote the joined text at the seam ('...приглашение на [continuation start]') and check that the continuation starts directly with a grammatically fitting object. It must not repeat 'на' or 'приглашение', and it must not open with a capitalized restart or a hedge like 'Возможно'/'Кажется'. A continuation that begins right where the prefix stops is correct and is not a repetition. A broken case or a mismatched start is a major failure."}, {"id": "c2", "description": "Voice: the continuation keeps the author's casual, wry stance on spam as an analogy joke in conversational Russian. Lowercase, slang, fragments, '..' and missing commas are acceptable. Heavily penalize assistant-like writing: bookish connectives ('безусловно', 'по-видимому', 'что весьма показательно'), balanced polished clauses, and a tidy moral or explanatory closer such as 'поэтому важно фильтровать почту'. A rough but lively phrasing beats a correct formal one."}, {"id": "c3", "description": "Substance: the analogy should land as an actual joke with a concrete image. Examples are a specific event the 'invitation' is to, such as a cousin's wedding or a Herbalife presentation, or a twist showing that it keeps coming because nobody revoked it. Stock hyperbole or a vague, generic completion (e.g. 'на что-то ненужное') is weak, and naturalness is capped around 8/10 without something distinctive. Filler clauses that add nothing get no credit."}, {"id": "c4", "description": "Hygiene, a minor weight compared with seam, voice and substance: the answer has no links, no meta-commentary or explanation of the task, and no invented hashtags. The full tweet (prefix plus continuation) stays within 280 characters. A playful in-voice hashtag is only a small flaw and never outweighs seam or voice problems."}]}, "target_words": 10, "reference": "вечеринку, где только ты и куча рекламных ботиков! #ОтвечаюСпамом #ТехнологииПроникаютВЖизнь", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1292, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_14b80cbad8ea250d90b4e694", "id": "14b80cbad8ea250d90b4e694", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1999", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 11 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nя музыкальный тормоз у меня новый ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends on the adjective 'новый' and needs a masculine noun next, e.g. 'любимый трек', 'фаворит' or a named song/artist. The judge must quote the joined text at the seam ('...у меня новый ' + continuation start) before deciding. Pass if the continuation starts directly with a fitting noun or noun phrase that agrees with 'новый'. Fail if it repeats 'новый' or 'у меня', restarts with a capitalized sentence, or opens with a hedge such as 'Кажется,' or 'Пожалуй,'. A continuation that begins exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps the self-deprecating, casual lowercase tone of 'я музыкальный тормоз'. That tone means admitting to discovering music late, for example a song everyone else has known for ages. Slang, run-ons, '..', fragments and enthusiasm are fine. Heavily penalize assistant-style phrasing: bookish connectives ('безусловно', 'что показательно', 'по-видимому'), balanced polished clauses, and tidy moralizing or explanatory closers. A formal rewrite scores below a rough but lively one."}, {"id": "c3", "description": "Substance: Reward concrete specifics, such as a named artist, track or year it came out, a concrete action like playing it on repeat, or an actual joke about being late to it. Treat generic lines like 'новая любимая песня, слушаю постоянно' or stock hyperbole as weak, and cap the score around 8 when nothing distinctive is present. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (minor weight): There should be no links, no meta-commentary or explanation of the task, no invented hashtags, and the whole tweet must fit within 280 characters. A playful hashtag in the author's own voice is only a small flaw and must never outweigh seam or voice failures."}]}, "target_words": 11, "reference": "краш Бурито. наверное, все его уже забыли, а я только узнала.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1999, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_16f4997c030f9b7fa3cb7b0f", "id": "16f4997c030f9b7fa3cb7b0f", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1426", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 17 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nБро вы не представляете сегодня получил ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'получил ' (a verb with a trailing space), which needs a direct object or complement. The judge must quote the joined text at the seam, e.g. 'получил [continuation start]', before deciding. Pass if the continuation starts directly with what was received (a noun phrase, number or similar) and reads grammatically as one sentence. Fail if it re-types 'получил', restarts with a capitalized new sentence or a hedge, or leaves the verb without an object. A continuation that begins exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps the excited, informal 'Бро вы не представляете' register. Russian slang, lowercase, missing punctuation, run-ons, exclamations and emoji-like enthusiasm all count as fitting. Heavily penalize assistant-style writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), polished balanced clauses, tidy morals, and explanatory or moralizing closers. A rough but lively line beats a formal, careful rewrite."}, {"id": "c3", "description": "Substance: The continuation names a specific, concrete thing that was received and/or a vivid reaction or joke (a named item, an amount, a person, a funny detail). Generic hype such as 'что-то невероятное' or 'лучший подарок' with nothing distinctive caps naturalness at about 8/10. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted lowest): The output contains no links, meta-commentary, quotation of the task, or explanation, and the full tweet fits within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and never outweighs a failure of seam or voice."}]}, "target_words": 17, "reference": "фишинг-сообщение от банка! Как будто у них нет других дел кроме как разводить людей! #Фишинг #Лохотрон #ИнтернетБезопасность", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1426, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_192c4a4d352fd6fb3cd94f31", "id": "192c4a4d352fd6fb3cd94f31", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:618", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 12 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда чё-то не так, а ты ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text around \"а ты \" plus the first words of the continuation before deciding. The prefix ends after the subject \"ты\" with a trailing space. The continuation should start with a verb or predicate phrase that grammatically completes \"ты ...\", for example a lowercase verb such as \"не понимаешь\". Penalize a continuation that re-types \"ты\" or \"а ты\", opens with a capitalized restart, or opens with a sentence-initial hedge such as \"Возможно\" or \"Кажется\". A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation should keep the casual, colloquial Russian register signalled by \"чё-то\". Lowercase, slang, fragments, '..' run-ons, nonstandard spelling and emotional exclamations are fine when they fit. Heavily penalize assistant-style writing: bookish hedges and connectives (\"возможно\", \"по-видимому\", \"безусловно\", \"что показательно\"), balanced polished clauses, and a tidy summarizing moral or explanatory closer about feelings or life. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific relatable situation, concrete action, named thing, or actual punchline that completes the 'when something's off and you...' setup, for example a particular reaction or a recognizable everyday image. Treat generic phrasing or stock hyperbole (\"просто ужасно\", \"это бесит\") with nothing distinctive as weak, capping naturalness around 8. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than seam, voice and substance): The output should contain no links, no meta-commentary or explanation of the task, and no quotation of the prefix, and the full tweet must stay under 280 characters. A playful hashtag in the author's voice is only a minor flaw. It must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "просто банишь всех подряд, как будто это новая диета! #Банить #ЖизньВТвиттере #СобериСвоюКоманду.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 618, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_19d3fd238f505e38e04cbb4e", "id": "19d3fd238f505e38e04cbb4e", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:93", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 10 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЯ оценил городскую новинку «Москвичи могут ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"…«Москвичи могут \" + continuation). The prefix ends after the modal verb \"могут\" inside an open quotation of a city initiative's name. The continuation must start directly with what Muscovites can do, for example an infinitive or verb phrase, and must not retype \"могут\" or \"Москвичи\". It must not open with a capitalized restart or a hedge. A continuation that starts exactly where the prefix stops is correct. The quoted slogan should plausibly close with » and then move to the author's evaluation."}, {"id": "c2", "description": "Voice and stance: The author speaks in first person (\"Я оценил\") about a city novelty. The continuation should give a casual, opinionated verdict, such as ironic, delighted or grumpy, in conversational Russian. Slang, lowercase, '..', fragments and emoticons are acceptable. Penalize heavily any assistant-like hedges or bookish connectives (\"возможно\", \"безусловно\", \"что весьма показательно\"), balanced polished clauses, and tidy moralizing or explanatory closers."}, {"id": "c3", "description": "Substance: Reward a concrete, plausible city service or action inside the slogan, such as paying for something via an app or getting something at a metro station. Also reward a specific reaction, detail or actual joke about it. Score generic praise or stock hyperbole (\"просто супер\", \"очень удобно\") with nothing distinctive as weak, capped around 8 out of 10. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight): Check that there are no links, no meta-commentary or explanation of the task, no repetition of the prompt, and that the full tweet stays within 280 characters. A playful, author-voiced hashtag is only a minor flaw and must never outweigh seam or voice failures."}]}, "target_words": 10, "reference": "присоединиться к акции «Волонтеры Нового года»» в проекте «Активный гражданин»", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 93, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_1f4502bd8cf3f3aa1b4836b9", "id": "1f4502bd8cf3f3aa1b4836b9", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:992", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 11 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКто-то сказал кьют? 😍💖 Все что ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam ('...Все что ' + continuation). The prefix ends on 'Все что ' with a trailing space, so the continuation must continue that clause directly, e.g. with a verb or phrase such as 'связано с котиками...'. It must not re-type 'Все что' or 'что', and it must not start a new capitalized sentence or a hedge. A continuation that picks up exactly at this point is correct and must not be marked as repetition. The resulting clause must read as grammatical, colloquial Russian."}, {"id": "c2", "description": "Voice: Keeps the gushing, informal, emoji-heavy enthusiasm of 'Кто-то сказал кьют? 😍💖'. Lowercase, slang, fragments, '..', exclamations and emoji are fine. Heavily penalize assistant-like register: bookish connectives or hedges (e.g. 'пожалуй', 'безусловно', 'что весьма показательно'), tidy morals, balanced polished clauses, or explanatory or moralizing closers. A rough but lively line beats a formal one."}, {"id": "c3", "description": "Substance: Reward a specific cute thing being named or shown, a concrete action, or an actual joke, for example particular animals, a person, an object, or a fandom reaction. Generic stock gushing alone (e.g. 'всё что милое — моё') is weak and capped near 8/10 for naturalness. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Contains no links, no meta-commentary, and no explanation of the task. The full tweet stays within 280 characters. An invented hashtag is a minor flaw, even one in a playful author voice, and must never outweigh a seam or voice failure."}]}, "target_words": 11, "reference": "меня окружает, будто превращается в милейшую панду! 🐼📸 #Кьют #Милота #НежностьНаМаксималках", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 992, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_213b2867f185541de724faa5", "id": "213b2867f185541de724faa5", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:3027", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 15 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nИз-за того что я сегодня жёг ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'Из-за того что я сегодня жёг ' (a verb awaiting an object or complement, followed by a space). The judge must quote the joined text at the seam, e.g. '...сегодня жёг [continuation start]', before deciding. Pass if the continuation starts directly with what was burned or how (e.g. an object like 'свечи', 'костёр', 'на тусе', a slang sense like 'жёг на танцполе') and then completes the 'из-за того что...' causal structure with a consequence. Fail if it re-types 'жёг' or earlier words, opens with a capitalized restart or a hedge ('Возможно', 'Кстати'), or never delivers the consequence clause. A continuation that begins exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual first-person Russian tweet register. Lowercase, slang, typos, missing commas, '..' run-ons, fragments, smileys like ')))' or exclamations are fine when they fit. Penalize heavily any assistant-like writing: hedges and bookish connectives ('по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, tidy morals, or explanatory/moralizing closers ('вот так и учишься осторожности'). A rough but lively line outranks a polished, hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific consequence or image tied to the burning, such as a smoky-smelling jacket, a soot-black face, a neighbor complaining, a hoarse voice after partying, or an actual joke or punchline. Generic phrasing or stock hyperbole ('теперь всё плохо', 'это было нечто') with nothing distinctive caps naturalness at about 8/10. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation. It has no links, no meta-commentary or explanation of the task, no invented hashtags, and the full tweet stays within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 15, "reference": "мусор Такой неприятный запах, но при том есть в нём какие-то нотки, что мне нравятся.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 3027, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_27f89d32a1cecc19c1c32a1c", "id": "27f89d32a1cecc19c1c32a1c", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1220", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 27 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\n💥 Ресёрч в этом году это ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The judge must quote the joined text at the seam (\"...Ресёрч в этом году это \" + the opening words of the continuation) before deciding. The prefix ends right after \"это\", so a predicate or complement is expected, such as a noun phrase, a comparison, or an evaluative fragment. Pass if the continuation supplies that complement directly and reads grammatically as one sentence. Fail if it repeats \"это\" or other prefix words, restarts with a capitalized new sentence, or opens with a hedge. A continuation that simply begins where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, punchy Russian tweet register, which the 💥 opener signals as an emphatic hot take about research this year. Slang, lowercase, run-ons with '..', fragments, exclamations and anglicisms are acceptable. Heavily penalize assistant-like writing: hedges and bookish connectives (\"пожалуй\", \"по-видимому\", \"безусловно\", \"что показательно\"), balanced polished clauses, tidy summarizing morals, and explanatory or moralizing closers. A rough but lively take beats a smooth, formal one."}, {"id": "c3", "description": "Substance: Reward a specific image, a named thing (a tool, field, trend, or concrete situation in research), a concrete action, or an actual joke or punchline that pays off the 'это...' setup. Generic hyperbole (\"это просто жесть\", \"это что-то невероятное\") or vague phrasing without anything distinctive is weak; cap naturalness around 8 in that case. Give no credit for filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check that the output contains only the continuation, with no links, meta-commentary, or explanation of the task. Check that the full tweet (prefix plus continuation) stays within 280 characters. A playful hashtag that fits the author's voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 27, "reference": "просто ездить по таинственным мира интернета, как по суперспособностям! Каждый день открываю что-то новенькое, и иногда даже удивляюсь, как сильно я в этом запутался! #НаучныйПутешественник #КтоТакойРесёрчер 🔍✨", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1220, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_2c5a13f0b4306a01ac32682e", "id": "2c5a13f0b4306a01ac32682e", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1279", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 16 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nНа минутку, вот и снюс пролетел ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends after the complete word 'пролетел' and a trailing space. The judge must quote the joined text at the seam, e.g. '...снюс пролетел [continuation start]', before deciding. A continuation that starts with a fitting complement such as 'мимо...', 'как...', a comma clause or a new thought is correct. It must not repeat 'пролетел' or 'снюс', and it must not open with a capitalized formal restart or a hedge like 'Возможно,'. A continuation that picks up exactly where the prefix stops must not be marked as repeating. Low marks go to a seam that is ungrammatical or semantically disconnected from the idea of snus having 'flown by'."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, slightly ironic Russian register, which may be wry, mock-dramatic or resigned. Slang, lowercase, run-ons, '..', fragments and colloquial forms like 'ну', 'блин' or 'короче' are fine where they fit. Heavily penalize assistant-like writing. This includes bookish connectives ('по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, and tidy morals or explanatory closers, for example health warnings about nicotine or remarks on how time flies. A rough but lively line scores above a correct but formal one."}, {"id": "c3", "description": "Substance: Reward a concrete detail, action, named thing or actual joke tied to the snus moment. Examples include what was missed, who it went past, a ban, a price, a friend, or the feeling afterward. Generic hyperbole or filler such as 'вот так вот бывает' with nothing specific caps the quality at a middling score. Padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer contains only the continuation. It has no links, no meta-commentary, no explanation of the task, and no quotation of the prompt. The full tweet stays within 280 characters. An invented hashtag that sounds like the author's playful voice is only a minor flaw and must not outweigh voice or seam failures."}]}, "target_words": 16, "reference": "мимо – как всегда, пробежал по шапке и унес все мои планы на вечер! 😂💨 #Снюсся #НеМогуСтопСнюс", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1279, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_330cdfd3632da00a2173dc40", "id": "330cdfd3632da00a2173dc40", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:631", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 18 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nБез б меня не остановляет никто! ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends after a complete exclamation ('Без б меня не остановляет никто! ') with a trailing space. The judge must quote the joined text at the seam before deciding. The continuation should start as a new thought or sentence that follows naturally from this boastful claim. It must not re-type 'никто!' or any other part of the prefix. It must not open with a hedge or a formal connective such as 'Впрочем', 'Возможно' or 'Конечно же,'. A capitalized new sentence is acceptable here because the prefix ended on '!'. A continuation that begins cleanly after the exclamation is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps the author's cocky, playful, colloquial Russian register. Slang, nonstandard spelling like 'остановляет', lowercase, fragments, extra exclamation marks and '..' run-ons are all acceptable. Heavily penalize assistant-style writing: hedges and bookish connectives (equivalents of 'perhaps', 'apparently', 'certainly', 'which is telling'), tidy morals, balanced polished clauses, and explanatory or moralizing closers that reflect on the boast. A rough but lively line beats a smooth, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete detail that grounds or extends the boast. Examples include a named thing, a specific action or plan, a situation, or an actual joke, such as who 'Б' might be or what the author will now do unchecked. Stock hyperbole or generic filler like 'я крутой, всё смогу' with nothing distinctive caps naturalness at about 8 out of 10. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains only the continuation. It has no links, no meta-commentary, and no explanation or translation. The full tweet stays within 280 characters. Invented hashtags are a minor flaw if they sound like the author's own playful voice, and must never outweigh a failure of voice or seam."}]}, "target_words": 18, "reference": "Жизнь это бесконечный поток возможностей, и я собираюсь штурмовать их, даже если придется лететь без б. #LifeGoals #NoLimits", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 631, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_34b3f5be0a9babe466e57057", "id": "34b3f5be0a9babe466e57057", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:620", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 14 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда ты просыпаешься и понимаешь что ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam ('...понимаешь что' + continuation) before judging. The prefix ends on the conjunction 'что', so the continuation must supply the clause it introduces. That means a subject or predicate describing what the person realizes, e.g. 'сегодня суббота' or 'проспал'. It must not repeat 'что' or 'понимаешь', and it must not restart with a capitalized new sentence or an introductory hedge like 'Возможно,'. A continuation that starts directly with the content of the realization is correct and must not be marked as repeating. Odd spacing in the prefix comes from corpus tokenization and should be ignored."}, {"id": "c2", "description": "Voice: The continuation reads like a casual Russian tweet in the 'когда ты просыпаешься и понимаешь' meme format. Lowercase, slang, missing punctuation, ')))' or '..', fragments and exclamations are acceptable. Heavily penalize assistant-like writing: bookish connectives or hedges (e.g. 'по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, and a tidy moral or explanatory closer. A rough, lively punchline beats a correct but formal one."}, {"id": "c3", "description": "Substance: The realization is specific, such as a named day, an exam, a forgotten alarm, a concrete situation or object, or a real comic twist or emotional punch. Stock phrasing like 'жизнь боль' or 'всё плохо' with nothing distinctive caps naturalness around 8/10. Filler clauses add no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): Check for links, meta-commentary, explanations of the task, and quotation of the prompt. The whole tweet must fit within 280 characters. A playful hashtag in the author's voice is only a minor flaw and must not outweigh seam or voice failures."}]}, "target_words": 14, "reference": "сегодня выходной и можно просто лежать на диване. Вот он, Baldyosh! #жизньпрекрасна #безделье 😍✨", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 620, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_3508099bafb13fa733f33ef2", "id": "3508099bafb13fa733f33ef2", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:989", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 24 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда твой дружище окунулся в мир ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends on 'окунулся в мир ' and needs a genitive noun or noun phrase (e.g. 'мир аниме', 'мир криптовалют'). The judge must quote the joined text at the seam, from 'в мир' through the first words of the continuation. Pass if the continuation starts directly with a fitting grammatically agreeing complement. Do not penalize it as repetition. Fail if it repeats 'мир' or 'в мир', starts with a capitalized restart or a hedge ('Возможно', 'Кажется'), or breaks case agreement."}, {"id": "c2", "description": "Coherence of the 'Когда...' frame: The tweet is a 'когда твой дружище…' meme setup. The continuation should finish the topic, then deliver the reaction or punchline, such as what happens to the friend or how the speaker feels. Examples: he now only talks about it, the speaker is lost, the speaker is annoyed or amused. Penalize if the setup never pays off or if it drifts into an unrelated statement."}, {"id": "c3", "description": "Voice: Keep a casual Russian Twitter register that matches 'дружище'. Slang, lowercase, fragments, '..', ')))' and emoji-like exclamations are fine. Heavily penalize assistant-like writing: bookish connectives ('очевидно', 'что весьма показательно', 'безусловно'), balanced polished clauses, and moralizing or summarizing closers (e.g. 'главное — поддерживать друзей'). A rough but lively line beats a tidy, formal one."}, {"id": "c4", "description": "Substance and hygiene: Reward a specific named world or hobby and a concrete detail or real joke, such as a quoted phrase the friend now says, a specific behavior, or a named game or show. Stock hyperbole or generic phrasing with nothing distinctive caps naturalness at about 8/10. Minor checks: no links, no meta-commentary or task explanation, and the whole tweet fits 280 characters. A playful in-voice hashtag is only a small flaw. This criterion weighs less than seam, coherence and voice."}]}, "target_words": 24, "reference": "куколд-отношений и теперь каждый вечер заказывает пиццу с другой а ты остаешься в одиночестве, как будто смотришь 100 сезонов драмы онлайн. 😂🍕 #куколд #друзья", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 989, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_38b4e43485ffcd637fae77bd", "id": "38b4e43485ffcd637fae77bd", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:882", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 12 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда кто-то говорит, что интернет это ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'интернет это ' (trailing space after 'это'), so a predicate noun or phrase is expected next. The judge must quote the joined text at the seam, e.g. '...интернет это [first words]'. Pass if the continuation starts directly with the predicate (e.g. 'зло', 'помойка', 'для школьников') and does not re-type 'это' or 'интернет'. Fail if it opens with a capitalized restart, a sentence-initial hedge (e.g. 'Возможно,', 'Честно говоря,'), or leaves the 'Когда кто-то говорит...' clause without a main-clause reaction. A continuation that begins exactly where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation must keep a casual Russian tweet register and the author's implied stance, an ironic or exasperated reaction to someone's claim about the internet (e.g. 'я просто молча закрываю вкладку', 'хочется спросить, откуда он это узнал'). Slang, lowercase, missing commas, '..', ')))' and fragments are legitimate. Penalize heavily: bookish connectives or hedges ('пожалуй', 'по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, and tidy moralizing closers such as 'ведь всё зависит от того, как его использовать'. A rough, lively reaction beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific reaction or joke. Examples include a named site or app, a vivid image, a pointed irony such as noting the person said it online, or a specific action the author takes. Generic stock phrasing or hyperbole without anything distinctive ('бесит', 'это просто ужас') counts as weak and caps naturalness around 8/10. Give no credit for filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Fail for any link, meta-commentary, explanation of the task, or quoting of the instructions. Check that the full tweet fits in 280 characters. Invented hashtags that sound like the author's own playful voice are only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 12, "reference": "не ИРЛ, я всегда отвечаю: А ты посмотри, сколько мемов мы создали!", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 882, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_39c41746b74619cf7da89044", "id": "39c41746b74619cf7da89044", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:2690", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 15 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nУ нас с майором самые лучшие ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends on the adjective 'лучшие' (plural), which needs a noun. Before deciding, the judge must quote the joined text at the seam, e.g. 'самые лучшие [first words]'. Pass only if the continuation starts directly with a fitting plural noun or noun phrase that agrees grammatically, such as 'отношения', 'тренировки' or 'планы'. Fail if it repeats 'лучшие' or 'самые', opens with a capitalized restart, or opens with a sentence-initial hedge. A continuation that begins exactly where the prefix stops is correct and is not repetition."}, {"id": "c2", "description": "Voice: The continuation keeps a casual, warm or boastful Russian tweet register about the author and 'майор', whether that is a friend, a nickname, a pet or a superior. Lowercase, slang, '..' run-ons, fragments, emoticons and exclamations are fine. Heavily penalize assistant-style Russian: hedges and bookish connectives ('пожалуй', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, and tidy moralizing or summarizing closers. A rough but lively line beats a formal one."}, {"id": "c3", "description": "Substance: Reward a concrete detail, such as a specific shared activity, place, named thing or in-joke, or an actual punchline about the duo. Generic phrasing like 'самые лучшие отношения в мире' with nothing distinctive is weak; cap naturalness around 8 out of 10. Give no credit for filler clauses."}, {"id": "c4", "description": "Hygiene (minor weight): The continuation contains no links, no meta-commentary and no explanation of the task, and the full tweet stays within 280 characters. A playful hashtag in the author's voice is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 15, "reference": "хэдканоны на переписку громоволков которые к сожалению, в адекватном виде даже в фанфик не перенести", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2690, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_3cf853caa05c06070e0c162d", "id": "3cf853caa05c06070e0c162d", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:749", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 19 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЗавтра иду на заброшку чтобы оставить ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"...чтобы оставить\" + continuation) before deciding. The prefix ends on the verb \"оставить\", which needs a direct object. The continuation must supply that object or its noun phrase directly, in the correct case (accusative), for example \"там свою...\". It must not retype \"оставить\", restart with a capitalized new sentence, or open with a hedge. A continuation that begins exactly where the prefix stops is correct and must not be marked as repeating. Fail it if the joined text is ungrammatical or leaves the object missing."}, {"id": "c2", "description": "Voice: The continuation keeps the casual first-person Russian tweet register: spoken phrasing, lowercase, missing commas, slang, '..' or ')))' are all fine. It keeps the author's stance as someone actually going to the abandoned building on purpose. Penalize heavily any assistant-like writing: bookish connectives or hedges (\"возможно\", \"по-видимому\", \"что весьма показательно\"), balanced polished clauses, a tidy moral, or a warning or moralizing closer (for example about safety or legality). A rough, lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete thing being left or done, such as a named object, a graffiti text, a message to someone, or a weird action, or a real punchline or twist that fits urbex culture. Generic filler or stock hyperbole without a distinctive detail caps the score around 8/10. Do not credit padding clauses that add nothing."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): no links, no meta-commentary or explanation of the task, no invented formal hashtags, and the full tweet stays within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 19, "reference": "там свое граффити! Все говорят, что это искусство, но меня больше волнует, что соседка все равно не поймет! #УличноеИскусство #Граффити", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 749, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_4182797f982bd310d625d65f", "id": "4182797f982bd310d625d65f", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:983", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 10 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКриптовалюта - это как шампанское для ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends on the preposition 'для' followed by a space, so it needs a noun phrase in the genitive case (e.g. 'для ...'). The judge must quote the joined text at the seam ('...шампанское для [continuation start]'). Pass if the continuation opens directly with a grammatically fitting genitive noun or noun phrase. Fail if it repeats 'для' or earlier prefix words, restarts with a capitalized new sentence, or opens with a hedge. A continuation that starts exactly at the noun must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps a casual, opinionated Russian tweet register and completes the simile with the author's ironic or punchy stance. Slang, lowercase, run-ons, '..', exclamations and nonstandard spelling are acceptable. Heavily penalize assistant-style writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, tidy morals, or explanatory or moralizing closers. A rough but lively line beats a formal hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, witty payoff for the champagne comparison, such as a concrete group of people, a recognizable image (hangover, bubbles, popping cork, New Year's), or an actual joke linking crypto and champagne. Generic phrasing or stock hyperbole without a distinctive detail caps naturalness at about 8. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (minor weight): The continuation has no links, no meta-commentary or task explanation, and the full tweet fits within 280 characters. Invented hashtags are a minor flaw only if they sound like the author's own playful voice. They must never outweigh seam or voice failures."}]}, "target_words": 10, "reference": "гиков! Кто еще по хорошей цене набирает битки? #Криптожизнь #ХочуБогатеть", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 983, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_46231ebd8fb41cdcfb0e3cba", "id": "46231ebd8fb41cdcfb0e3cba", "domain": "tweets", "language": "ru", "split": "validation", "group": "ru:rulexnorm:746", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 12 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда гопники на районе делают утреннюю ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'делают утреннюю ' (a feminine accusative adjective that still needs a noun). The judge must quote the joined text at the seam, e.g. '...делают утреннюю [continuation]', before deciding. Pass if the continuation starts directly with a noun or noun phrase that agrees grammatically ('зарядку', 'пробежку', 'разминку', etc.) or another grammatical completion. Fail if it re-types 'утреннюю' or earlier words, opens with a capitalized restart, or opens with a hedge or new sentence that leaves the adjective dangling. A continuation that begins exactly where the prefix stops must not be marked as repetition."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, joking Russian tweet register about neighborhood gopniks. It may use slang, lowercase, run-ons, '..', fragments, exclamations or nonstandard spelling. Heavily penalize assistant-like writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что весьма показательно'), tidy morals, balanced polished clauses, and explanatory or moralizing closers. A rough but lively completion beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete, distinctive payoff such as a specific gopnik image (корточки, семки, адидас, пиво, турник у гаражей), a concrete action, or an actual punchline that completes the 'когда...' setup. Stock hyperbole or a generic reaction with no distinctive detail is weak, and its naturalness is capped around 8/10. Give no credit for filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation has no links, no meta-commentary or task explanation, no invented informational hashtags, and the full tweet stays within 280 characters. A hashtag that playfully fits the author's voice is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 12, "reference": "зарядку, я понимаю, что сможет даже спортзал им завидовать. 💪😆 #ГопникНаСпорте #УтреннийШум", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 746, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_499d2e32aeb5ef36e4765ce8", "id": "499d2e32aeb5ef36e4765ce8", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:3690", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 16 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКстати говоря я при всём при ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'при всём при ' (a trailing space after the preposition). This sets up the idiom 'при всём при том'. Quote the joined text at the seam before deciding. A continuation that begins with 'том', or with another word that completes the phrase grammatically, is correct. Penalize a continuation that re-types 'при' or 'всём', starts with a capitalized restart or a fresh sentence, or opens with a hedge. Do not mark a continuation as repeating when it starts exactly where the prefix stops."}, {"id": "c2", "description": "Voice: The text should keep a casual, first-person, chatty Russian register in which the speaker is about to confess or assert something. Examples of this register are 'кстати говоря' and 'я при всём при том…'. Lowercase, slang, missing commas, '..', fragments and emotional exclamations are acceptable. Heavily penalize assistant-like writing, including bookish connectives and hedges such as 'вероятно', 'безусловно' and 'что показательно', balanced polished clauses, and a moralizing or summarizing closer."}, {"id": "c3", "description": "Substance: Reward a continuation that delivers a concrete personal admission, opinion, named thing, action or joke that pays off the 'при всём при том' setup, for example a specific contrast or a quirky preference. Generic statements, stock hyperbole or filler clauses with nothing distinctive are weak and should cap the naturalness score at about 8/10."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer must contain no links, no meta-commentary or explanation of the task, and no invented hashtags, and the full tweet must stay within 280 characters. A playful hashtag in the author's own voice is only a minor flaw. It should never outweigh a failure of seam or voice."}]}, "target_words": 16, "reference": "этом ещё хочу написать жутко дурацкую смску одному товарищу М. Слушайте я основательно вернула свои 16!", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 3690, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_69eedcc78decb85e81e9a380", "id": "69eedcc78decb85e81e9a380", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:684", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 11 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nВчера на вечеринке такой крутой вайб ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam ('...такой крутой вайб' + continuation) before judging. The prefix ends on a complete noun phrase followed by a space, so the continuation should flow on naturally. Good openings include a clause like 'был, ...', a dash, an exclamation, or a new lowercase/colloquial fragment. Penalize if it re-types 'вайб' or other prefix words, or opens with a formal capitalized restart or hedge (e.g., 'Возможно,', 'Безусловно'). A continuation that starts exactly after the prefix must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's enthusiastic, casual Russian party-talk register and positive stance toward the party. Slang, lowercase, run-ons with '..', fragments, exclamations and nonstandard spelling are acceptable. Heavily penalize assistant-like writing: bookish connectives ('что говорит о', 'по-видимому', 'безусловно'), balanced polished clauses, and tidy morals or reflective summaries ('такие моменты важно ценить'). A rough, lively line beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward concrete party detail, such as a specific song or track, a person or friend doing something, dancing until morning, a named place or drink, or an actual joke. Generic hype ('было очень круто, всем понравилось', 'незабываемо') with nothing distinctive should score at most about 8/10 for naturalness. Do not credit filler clauses that add no image or action."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output contains only the continuation. Penalize links, meta-commentary or task explanation, English text, and a full tweet exceeding 280 characters. Invented hashtags are a minor flaw if they sound like the author's playful voice, and must never outweigh voice or seam failures."}]}, "target_words": 11, "reference": "был что я даже забыла, как зовут свою собаку! #вечеринка #вайб🏄‍♂️", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 684, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_6aa6802ab2d8c32787d38822", "id": "6aa6802ab2d8c32787d38822", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1261", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда твой друг скинул тебе ссылки ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"...скинул тебе ссылки\" + continuation). The prefix ends after the noun \"ссылки\" with a trailing space, so the continuation should naturally extend the clause. Examples are \"на...\", \"на свои любимые...\", or a comma leading to the main clause of the \"Когда...\" setup. It must not re-type \"ссылки\" or restart with a capitalized new sentence or hedge. The joined sentence must read grammatically and coherently as Russian, and the \"Когда\" setup should get a payoff, such as a reaction or punchline. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps a casual Russian tweet register. Lowercase, slang, run-ons, \"))\" or emoji-like enthusiasm, and nonstandard spelling are fine. Heavily penalize assistant-style writing: hedges or bookish connectives (\"возможно\", \"по-видимому\", \"безусловно\", \"что говорит о многом\"), tidy morals, balanced polished clauses, and explanatory closers. A rough but lively reaction beats a formal or hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, relatable detail or an actual joke about the links the friend sent. Examples are naming what the links are (мемы, видео с котами, порно-спам, статьи), a concrete reaction, or a punchline about friendship. Generic filler like \"и это так мило\" or stock hyperbole with nothing distinctive caps the score around 8/10. Padding clauses get no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check for actual URLs, invented hashtags, meta-commentary or explanation of the task, non-Russian output, and a combined tweet over 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw. It should never outweigh a voice or seam failure."}]}, "target_words": 13, "reference": "на акции, а ты понял, что это просто скамеры в действии! #НеПопадисьНаСкачки #СкамерОЧЕНЬОСТОРОЖНО", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1261, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_77d3e78e2a8ff8081650e16c", "id": "77d3e78e2a8ff8081650e16c", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:877", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 12 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nИмхо если бы все просто слушали ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with 'Имхо если бы все просто слушали ' (after a verb that needs an object, with a trailing space). The judge must quote the joined text at the seam, e.g. '...просто слушали [continuation start]', before deciding. The continuation should begin directly with what is being listened to, such as an object in the accusative or 'друг друга'. It should then complete the conditional with a consequence clause, typically 'то...' or an equivalent. Penalize three things: re-typing 'слушали' or other prefix words, a capitalized restart, and a sentence-initial hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, opinionated 'имхо' register. Lowercase, slang, missing punctuation, '..' run-ons and fragments are fine. Penalize assistant-style writing heavily. This includes hedges and bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что весьма показательно'). It also includes balanced, polished clauses and a tidy moralizing or explanatory closer (e.g. 'ведь умение слушать — ключ к взаимопониманию'). A rough but lively take is better than a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific object of listening, a concrete consequence, a named thing, or an actual joke or ironic twist, such as a particular band, a person, a mom's advice, or a specific absurd outcome. Treat generic, stock completions as weak and cap them around 8 when nothing distinctive accompanies them; examples are 'друг друга, мир был бы лучше' or 'то всё было бы хорошо'. Give no credit for filler clauses or padding."}, {"id": "c4", "description": "Hygiene (lower weight): Check for links, meta-commentary, task explanations, and quotation of the prefix. Also check that the full tweet stays within 280 characters. An author-like playful hashtag is only a minor flaw and must not outweigh seam or voice failures."}]}, "target_words": 12, "reference": "друг друга, мир стал бы местом фейерверков и счастья! #мирдолжендружить #слушайте #имхо", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 877, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_7969b341dffa7b974f9f70fb", "id": "7969b341dffa7b974f9f70fb", "domain": "tweets", "language": "ru", "split": "validation", "group": "ru:rulexnorm:2523", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 15 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nвау ложусь спать в пять утра ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The judge quotes the joined text at the seam (\"...ложусь спать в пять утра \" + continuation) before deciding. The prefix ends after a complete phrase with a trailing space, so the continuation should flow on naturally. Good flow means adding to the thought (e.g., a reason, a consequence, or an ironic remark about waking up) in the same lowercase run-on style. It must not re-type \"утра\" or restart with a capitalized formal sentence or a hedge like \"Возможно\" or \"Кажется\". A continuation that starts right where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps the author's ironic, informal stance. The \"вау\" is self-mocking about going to bed at 5 a.m. Lowercase, slang, fragments, '..', and emoji-like enthusiasm are acceptable. Penalize heavily any assistant-style writing, such as bookish connectives (\"очевидно\", \"безусловно\", \"что показательно\"), balanced polished clauses, tidy morals, or moralizing closers (e.g., advice about sleep hygiene). A rough but lively continuation beats a hedged, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete detail or real joke. Examples are what kept the author up (a series, a game, scrolling, a deadline), a specific wake-up time or an alarm, or a funny consequence for tomorrow. Generic filler or stock hyperbole (\"это было ужасно\", \"жизнь боль\") with nothing distinctive caps the score around 8/10. Padding clauses get no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The continuation has no links, no meta-commentary or explanation of the task, and no invented formal hashtags, and the full tweet stays within 280 characters. A playful hashtag in the author's own voice is a minor flaw only. It must never outweigh a failure of voice or seam."}]}, "target_words": 15, "reference": "хотя вчера ужасно устала и хотела лечь «через пол часа» бляяять как научится контролировать себя", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2523, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_7a5c9fa6c65de4f2eb19a5fe", "id": "7a5c9fa6c65de4f2eb19a5fe", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:5229", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nУ женщины всего две обязанности перед ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"...две обязанности перед\" + continuation's first words) before deciding. The prefix ends on the preposition \"перед\", so the continuation must open with a noun phrase in the instrumental case that completes it (e.g. \"мужчиной\", \"обществом\", \"собой\"). It must not re-type \"перед\", restart with a capitalized sentence, or open with a hedge. Starting exactly at the noun is correct and is not repetition. Fail if the grammar at the seam breaks."}, {"id": "c2", "description": "Coherence and stance: The continuation should deliver on the setup \"всего две обязанности\". It should name or clearly imply the two duties, or twist the setup into a punchline, while keeping the author's position (ironic, provocative or sarcastic) consistent. Penalize drifting into an unrelated topic, leaving the 'two duties' frame unresolved, or neutralizing the stance into a balanced or apologetic take."}, {"id": "c3", "description": "Voice: The text should read as a colloquial Russian tweet. Lowercase, slang, fragments, '..' and exclamations are fine. Heavily penalize assistant-style writing: hedges and bookish connectives (\"пожалуй\", \"безусловно\", \"что показательно\", \"по-видимому\"), polished balanced clauses, and moralizing or explanatory closers or tidy summaries. A rough but lively line beats a formal, hedged one."}, {"id": "c4", "description": "Substance and hygiene: Reward a specific, concrete image or an actual joke, such as named duties or a sharp reversal. Stock phrasing or generic hyperbole with nothing distinctive caps the naturalness score around 8 out of 10. Check for links, meta-commentary, task explanations, and whether the full tweet exceeds 280 characters. Playful in-voice hashtags are only a minor flaw. This criterion weighs less than seam, coherence and voice."}]}, "target_words": 13, "reference": "мужчиной: 1. Успокоить его, когда он нервный. 2. Нервировать его, когда он спокойный.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 5229, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_83b0a33bc930fcfee0a3a055", "id": "83b0a33bc930fcfee0a3a055", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1513", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 17 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЧикибамбони это не просто слово, это ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'это ' (after a comma, mid-construction 'не просто слово, это ...'). Quote the joined text at the seam, e.g. '...это [first words]'. The continuation must directly supply what the word IS (a noun phrase or clause completing 'это ...'), without re-typing 'это', without restarting with a capitalized new sentence or a hedge like 'Пожалуй'/'Наверное'. A continuation that starts exactly after 'это' is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: Continuation keeps the playful, enthusiastic, colloquial Russian tweet register of someone hyping a silly word ('Чикибамбони'). Lowercase, slang, fragments, '..', exclamations, and nonstandard spelling are fine. Heavily penalize assistant-style writing: bookish connectives or hedges ('по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, tidy summarizing morals or explanatory/moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific, funny or vivid completion — a concrete image, situation, named thing or actual joke about what 'чикибамбони' means (e.g. a mood, lifestyle, state of mind with a concrete detail). Stock hyperbole like 'это целая философия/образ жизни' with nothing distinctive caps naturalness around 8/10; filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, no meta-commentary or explanation of the task, no quotes/labels around the answer, and full tweet within 280 characters. Invented hashtags are a minor flaw if they sound like the author's playful voice; never let them outweigh seam or voice failures."}]}, "target_words": 17, "reference": "состояние души! Когда жизнь дает лимоны, я превращаю их в чикибамбони и танцую под дождем! #Позитив #ЧикибамбониВперед", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1513, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_8b4c7803c8414b547b8aab3f", "id": "8b4c7803c8414b547b8aab3f", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:1202", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 15 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЕсли кто-то скажет, что я не ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'что я не ' (a negation awaiting a verb, adjective or noun). The judge should quote the joined text at the seam, e.g. '...что я не [first words]', and confirm it reads as one grammatical Russian sentence. The continuation must start directly with what follows 'не'. Fail it if it re-types 'не' or 'я не', restarts with a capitalized new sentence, or opens with a hedge like 'Возможно' or 'Наверное'. A continuation that begins exactly after 'не' must not be marked as repeating."}, {"id": "c2", "description": "Coherence of the conditional: The 'Если кто-то скажет, что я не X' setup should get a fitting payoff in the author's voice, such as a retort, a threat, an ironic denial or a self-deprecating twist. Penalize continuations that leave the conditional logically dangling or contradict the setup."}, {"id": "c3", "description": "Voice: The continuation should keep a casual Russian tweet register. Slang, lowercase, '..', emoji-like enthusiasm, swearing-lite and fragments are acceptable when they fit. Heavily penalize assistant-like writing: bookish connectives or hedges ('пожалуй', 'безусловно', 'что весьма показательно'), balanced polished clauses, and tidy morals or explanatory closers. A rough but lively reply beats a formal, hedged one."}, {"id": "c4", "description": "Substance and hygiene: Reward a concrete image, named thing, specific action or real joke, and treat stock hyperbole or generic filler as weak (cap around 8/10 if nothing distinctive). Check that there are no links, no meta-commentary or task explanation, and that the full tweet stays within 280 characters. Invented hashtags are a minor flaw and must not outweigh voice or seam failures."}]}, "target_words": 15, "reference": "умею варить кофе, просто покажите им мой пруф фото с идеальной карамелизацией! #кофемания #бариста #пруф", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1202, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_8d903e61c2a8ec55a5d774c4", "id": "8d903e61c2a8ec55a5d774c4", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:542", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 25 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда пытаешься понять, что такое айпи, ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The judge must first quote the joined text at the seam (\"...что такое айпи, \" + the first words of the continuation). The prefix ends after a comma, inside a subordinate 'Когда...' clause, so the continuation should supply the main clause or continue the thought naturally, e.g. a consequence or reaction. Penalize re-typing 'айпи' or 'Когда пытаешься', and penalize a capitalized restart or a new standalone sentence that ignores the dangling 'Когда' clause. A continuation that picks up exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps a casual, self-ironic Russian tweet voice about the frustration or absurdity of trying to understand what an IP is. Lowercase, slang, fragments, '..', ')))' and nonstandard spelling are fine when they fit. Heavily penalize assistant-like writing: hedges and bookish connectives (например 'возможно', 'по-видимому', 'безусловно', 'что весьма показательно'), polished balanced clauses, tidy morals, and explanatory or educational closers such as actually defining an IP address. A rough but lively line beats a correct, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete image, a named thing (провайдер, роутер, цифры с точками, гугл, сисадмин), a specific action, or a real joke or punchline about the confusion. Stock hyperbole ('мозг взрывается', 'ничего не понятно') and generic phrasing with nothing distinctive alongside it cap the score around 8/10 on naturalness. Give no credit for filler or padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The answer contains no links, no meta-commentary, no notes about the task, and no quotation of the prefix. The whole tweet (prefix plus continuation) fits within 280 characters. Invented hashtags are a minor flaw if they sound like the author's playful voice, but they must never outweigh a failure of seam or voice."}]}, "target_words": 25, "reference": "и осознаешь, что это не просто набор цифр, а целый мир! А вы знали, что без айпи мы бы все заблудились в интернете? #айпи #интернетмагия", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 542, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_93ae7fdad31da8a1b4c16c71", "id": "93ae7fdad31da8a1b4c16c71", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:4660", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nнет наркотика хуже чем прогуливать серьёзно ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'серьёзно ' (an intensifier after the claim 'нет наркотика хуже чем прогуливать'). The judge must quote the joined text at the seam, e.g. '...прогуливать серьёзно [continuation]', before deciding. The continuation must read as a direct grammatical and logical extension. It should either carry on the thought (for example, why skipping is addictive or what happens once you start) or treat 'серьёзно' as an emphatic aside followed by more text. It must not re-type 'серьёзно' or other prefix words. It must not restart with a capitalized formal sentence or an opening hedge. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation must keep the author's informal, lowercase, jokey-confessional Russian register. It must also keep the stance that skipping classes is addictive like a drug. Slang, missing punctuation, run-ons, '..', fragments, exclamations and nonstandard spelling are acceptable. Heavily penalize assistant-style writing: bookish connectives or hedges ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, tidy morals, and moralizing or explanatory closers (for example, warnings that skipping is bad for your education). A rough, lively phrasing beats a correct, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific detail or an actual joke that develops the drug metaphor or the skipping experience. Examples: 'раз пропустил одну пару и всё', 'лежишь дома а совесть молчит', a named subject, or a ломка/доза image. Generic hyperbole or filler such as 'это реально ужасно' or 'очень затягивает' with nothing distinctive caps naturalness around 8/10. Padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer contains only the continuation. It has no links, no meta-commentary, no quotes around the output and no explanation of the task. The full tweet stays within 280 characters. A hashtag that fits the author's own playful voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 13, "reference": "я вчера никуда не пошла и знаете Я И СЕГОДНЯ НИКУДА НЕ ПОЙДУ", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 4660, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_9737a745d1ede7acb4efd924", "id": "9737a745d1ede7acb4efd924", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1373", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 12 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда ты понимаешь, что ты мач ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam (\"...что ты мач\" + continuation) before judging. The prefix ends right after \"мач\", which could be a truncated word (e.g. \"мачо\", \"мачете\", \"матч\") or the slang word \"мач\" (match). The continuation must join grammatically from that exact point. It may complete the word (e.g. \"о\" → \"мачо\") or add a fitting next word. It must not re-type \"мач\"/\"ты мач\" and must not restart with a capitalized sentence or a hedge. A continuation that starts exactly where the prefix stops is correct and is not repetition. Penalize heavily if the joined text is ungrammatical or nonsensical."}, {"id": "c2", "description": "Voice: The continuation keeps the conversational, self-ironic Russian tweet register of a \"Когда ты понимаешь, что...\" meme. Lowercase, slang, fragments, \"..\", run-ons, emoji-like exclamations and nonstandard spelling are all fine. Heavily penalize assistant-style writing: bookish connectives (\"пожалуй\", \"безусловно\", \"что весьма показательно\"), polished balanced clauses, explanatory or moralizing closers, and tidy summarizing morals. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: The line delivers a concrete punchline, situation, named thing or action that pays off the \"когда понимаешь\" setup, such as a specific funny realization. Stock hyperbole or generic phrasing (\"это так круто\", \"лучше всех\") with nothing distinctive caps the score at about 8/10 for naturalness. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The answer contains only the continuation, written in Russian. It has no links, no meta-commentary and no explanation of the task, and the full tweet stays within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and must not outweigh a voice or seam failure."}]}, "target_words": 12, "reference": "в своей области, но все еще сидишь в пижаме и ешь чипсы.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1373, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_975080df79cfdbf71bde2ca7", "id": "975080df79cfdbf71bde2ca7", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1175", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 18 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nПов-ка пов-ка где ж ты? В ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with the preposition 'В ' followed by a space. Quote the joined text at the seam ('...где ж ты? В [continuation start]') before deciding. The continuation must begin with a word that grammatically completes 'В' (a place or noun phrase in prepositional or accusative case, e.g. 'В холодильнике...'). It must not re-type 'В' or start a new capitalized sentence or hedge instead. A continuation that starts exactly at that point is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The continuation keeps the playful, colloquial, sing-song search for 'повидло/повка' (mock-lament, informal Russian). Lowercase, slang, fragments, '..', exclamations and nonstandard spelling are fine. Heavily penalize assistant-like writing: bookish connectives or hedges ('пожалуй', 'видимо', 'безусловно', 'что показательно'), balanced polished clauses, summarizing morals, or explanatory closers."}, {"id": "c3", "description": "Substance: Reward concrete, specific content, such as a named place where it might be (shelf, a grandma's jar, the dacha), a concrete action, or an actual joke or punchline. Generic filler or stock hyperbole without a distinctive detail is weak, and naturalness should be capped around 8 in that case. Do not credit padding clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check for links, meta-commentary, task explanations, or notes about the prompt, and confirm the full tweet stays within 280 characters. Invented hashtags are a minor flaw only if they sound like the author's playful voice. They must never outweigh seam or voice failures."}]}, "target_words": 18, "reference": "жизни каждого из нас наступает тот момент, когда пов стоит на первом месте в списке приоритетов! #ЖизненныеПриоритеты #ПовкаНаПервомМесте", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1175, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_9d81f210dcbc25bb639896d6", "id": "9d81f210dcbc25bb639896d6", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:2505", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 16 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nБл такой красивый парень снился ((9(\\n ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends after a complete phrase \"Бл такой красивый парень снился ((9(\" followed by a newline. The judge should quote the joined text at the seam before deciding. The continuation should read as a natural next line or thought. It must not re-type \"снился\" or other prefix words. It must not open with a formal restart or hedge (e.g. \"Возможно,\" \"Кстати говоря,\"). A continuation that simply starts the next line is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: keeps the author's wistful, slightly sad, casual register: a mild swear-abbreviation, sad brackets like \"((9(\", lowercase, fragments and run-ons are all welcome. Heavily penalize assistant-style output: bookish connectives (\"вероятно\", \"безусловно\", \"что весьма показательно\"), polished balanced clauses, tidy morals, and explanatory or moralizing closers such as reflections on what dreams mean. A rough, lively line beats a correct, formal one."}, {"id": "c3", "description": "Substance: reward a concrete detail or real joke, e.g. what happened in the dream, the disappointment of waking up, or a specific image of the guy or the situation. Generic filler like \"жаль что это был сон\" with nothing distinctive counts as weak and caps naturalness around 8. Padding clauses get no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): no links, no meta-commentary or task explanation, and the full tweet fits within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and should never outweigh a voice or seam failure."}]}, "target_words": 16, "reference": "Точеное лицо золотые кудри до лопаток зеленые глаза а взгляд то какой добрый и понимающий ах", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2505, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_9face34dab16d4d83ffa9246", "id": "9face34dab16d4d83ffa9246", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:2153", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nвпервые в жизни вышла в топике, ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends after a comma following 'впервые в жизни вышла в топике,'. The judge should quote the joined text at the seam, e.g. 'вышла в топике, [continuation]'. Then check that the continuation picks up naturally as the next clause, typically a reaction, a consequence or a weather/people detail. It should not repeat 'в топике' or other prefix words. It should not restart with a capital letter or a sentence-initial hedge like 'Возможно' or 'Кстати говоря'. Feminine verb agreement (вышла) must be kept if the author is referred to. A continuation that starts exactly at the comma is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice: The text should stay in the author's casual Russian tweet register. Lowercase, slang, colloquial forms, '..', ')))' style smileys, exclamations and fragments are all fine. Heavily penalize assistant-like phrasing: bookish connectives such as 'безусловно', 'по-видимому', 'что весьма показательно'; tidy morals about self-confidence or body positivity; balanced polished clauses; explanatory closers. A rough, lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete detail or joke tied to wearing a crop top for the first time. Examples include someone staring, freezing, a mom's reaction, a specific place or feeling exposed in the belly. Stock hyperbole or generic phrasing like 'это было незабываемо', with nothing distinctive alongside it, caps quality around 8/10. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output must contain no links, no meta-commentary or explanation of the task, and no Russian/English notes outside the tweet. The whole tweet must fit within 280 characters. A playful author-like hashtag is only a minor flaw and should never outweigh seam or voice failures."}]}, "target_words": 13, "reference": "причем с накидкой на плечи сверху\\n\\nпрошел какой-то мужик в компании и выставил руку…", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2153, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_a2a73ba8e411e9104c623423", "id": "a2a73ba8e411e9104c623423", "domain": "tweets", "language": "ru", "split": "train", "group": "ru:rulexnorm:2", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 14 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nник ты мог чуть позже сказать ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'ник ты мог чуть позже сказать ' (a verb expecting an object/clause, addressed to 'ник'). The judge must quote the joined text at the seam (e.g. '...позже сказать [continuation start]') before deciding. Pass if the continuation goes straight on from 'сказать' with what Nik could have said later, a 'что…' clause, or a reason/consequence, grammatically and without re-typing 'сказать' or other prefix words. Fail if it opens with a capitalized restart, a sentence-initial hedge, or a new unrelated sentence. A continuation that simply starts where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, lowercase, mildly exasperated or teasing reproach toward Nik, like a friend complaining about bad timing (spoiler, bad news, a plan change). Lowercase, slang, fragments, '..', ')))' and nonstandard spelling are fine. Heavily penalize assistant-style writing: bookish connectives or hedges ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, tidy morals, or explanatory or moralizing closers. A rough, lively complaint beats a formal, neat one."}, {"id": "c3", "description": "Substance: Reward a concrete detail of what was said or why the timing hurt, such as a named spoiler, a specific event like an exam, sleep or a match, a concrete consequence, or a real joke. Generic lines like 'было бы лучше' or stock hyperbole with nothing specific cap naturalness around 8/10. Padding or filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The continuation contains no links, meta-commentary, quotation of the task, or explanation, and the full tweet stays within 280 characters. An invented hashtag that fits the author's voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 14, "reference": "сабрине что время пришло мне спать надо блять но я хочу посмотоеть пятую серию", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_a425c62431426ed132a7e8e0", "id": "a425c62431426ed132a7e8e0", "domain": "tweets", "language": "ru", "split": "validation", "group": "ru:rulexnorm:1276", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nКогда ты пытаешься слить свои переживания ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends after 'Когда ты пытаешься слить свои переживания ' (trailing space, complete word, mid-clause, ending on a direct object). The judge must quote the joined text at the seam before deciding. The continuation should carry the 'Когда ты…' clause forward naturally. Typical ways to do that are adding a recipient or destination (e.g. 'подруге/в чат/в твиттер'), or completing the clause and delivering the main clause or punchline (often after a comma or dash). Penalize two kinds of restart. The first is re-typing 'переживания' or other prefix words. The second is opening with a capitalized new sentence or a hedge that abandons the 'Когда…' structure. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. The joined tweet should read as one coherent thought."}, {"id": "c2", "description": "Voice and stance: The continuation keeps a casual, relatable Russian Twitter register in a 'когда ты…' meme/situational style. Lowercase, slang, fragments, '..', ')))' or emoji-like enthusiasm, and nonstandard spelling are fine when they fit. Heavily penalize assistant-style writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), tidy summarizing morals, balanced polished clauses, and explanatory or moralizing closers about mental health. A rough but lively line beats a formal or hedged one."}, {"id": "c3", "description": "Substance: Reward a concrete payoff, such as a specific recipient or action, an actual ironic twist or joke, or a recognizable situation (e.g. the friend replies with a sticker, the person starts complaining about their own problems, or the message stays unread). Treat stock hyperbole or generic phrasing ('а всем всё равно', 'это так сложно') with nothing distinctive as weak, and cap its naturalness score around 8/10. Give no credit to padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output must contain no links, no meta-commentary or explanation of the task, and no invented hashtags. The full tweet must stay within 280 characters. A playful author-voiced hashtag is only a minor flaw and must never outweigh a failure of seam or voice."}]}, "target_words": 13, "reference": "в твит, а в итоге сливаешь свою жизнь в один поток🤪😂 #слив #жизнькакприключение", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1276, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_a95ad6d6d3025c01c343cd24", "id": "a95ad6d6d3025c01c343cd24", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1514", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 17 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nВсе эти разговоры о чечиках просто ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends with 'Все эти разговоры о чечиках просто ' — a dangling adverb 'просто' awaiting a predicate (e.g., a verb or adjective like 'бесят', 'достали', 'смешны'). The judge must quote the joined text at the seam (…'просто' + first words of the continuation) and check that it reads as one grammatical sentence. The continuation must not re-type 'просто' or other prefix words, and must not open with a capitalized restart or a sentence-initial hedge ('Возможно', 'Кажется'). A continuation that begins directly with the predicate is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: the continuation keeps the author's casual, irritated or dismissive Russian Twitter register about the chatter around 'чечики' (whatever the reaction—annoyance, mockery, fatigue). Slang, lowercase, '..', ')))', fragments and nonstandard spelling are acceptable. Heavily penalize assistant-like writing: bookish connectives or hedges ('по-видимому', 'безусловно', 'что весьма показательно'), balanced polished clauses, tidy morals, or explanatory or moralizing closers. A rough, lively reaction scores above a smooth, neutral rewrite."}, {"id": "c3", "description": "Substance: reward a concrete detail, such as a named situation, a specific action ('выключаю звук', 'листаю дальше'), a vivid image or a real joke. Generic hyperbole on its own ('просто ужас какой-то, сил нет') with nothing distinctive caps the score around 8/10. Filler clauses that add nothing earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): no links, no meta-commentary or explanation of the task, no quoting of the prefix, and the full tweet fits within 280 characters. A playful, author-like hashtag is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 17, "reference": "выбивают из колеи! Говорите что хотите, а я считаю, что настоящий чечик это искусство! #чечик #искусство #долгожданное", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1514, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_b0fb86adb14b9aa93142ae25", "id": "b0fb86adb14b9aa93142ae25", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:319", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 10 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nНе дружи с долбоебами постящими шутки ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends with 'постящими шутки ' (a noun followed by a space), so the continuation should naturally qualify or extend 'шутки' (e.g., 'про ...', 'из ...', 'которые ...') or move on to the next clause. The judge must quote the joined text at the seam before deciding. Penalize re-typing 'шутки' or 'постящими', or a capitalized restart or new sentence that ignores the dangling phrase. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: the continuation keeps the author's blunt, crude, advice-giving Russian Twitter register (mat like 'долбоебы' is fine, as are lowercase, slang, fragments, and nonstandard spelling) and keeps the stance of contempt for these people. Heavily penalize assistant-style writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, softened or moralizing explanations, and tidy summarizing morals. A rough but lively line beats a formal rewrite."}, {"id": "c3", "description": "Substance: reward a concrete, recognizable target or punchline, such as a specific kind of joke, meme, platform, or habit, or an actual witty kicker. Generic phrasing or stock hyperbole with nothing distinctive caps naturalness around 8 out of 10. Give no credit for padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): the output contains only the continuation, in Russian, with no links, meta-commentary, quotation marks, or explanation of the task, and the full tweet stays within 280 characters. Invented hashtags are a minor flaw unless they clearly fit the author's playful voice, and they never outweigh a seam or voice failure."}]}, "target_words": 10, "reference": "ургантов с киркоровыми:сам незаметно в такого обрротишься, то бишь-долбоебом станешь!", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 319, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_b38e2bce9b97ff9f0f1fa5de", "id": "b38e2bce9b97ff9f0f1fa5de", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1037", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 10 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЗаметила, что даже в сложные времена можно ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'можно ' (a modal requiring an infinitive or verb phrase). Quote the joined text at the seam (e.g. '...можно [first words of continuation]') before judging. The continuation must begin directly with an infinitive/verb phrase or a fitting complement that completes 'можно', without re-typing 'можно', without a capitalized restart or a new sentence, and without an opening hedge like 'Наверное,' or 'Конечно,'. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating. Seam breaks are a major failure."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, first-person, female ('заметила') observational register and her mildly positive, life-noticing stance. Informal phrasing, lowercase, fragments, '..', slang, emoji-like enthusiasm are fine. Heavily penalize assistant-like writing: hedges and bookish connectives ('пожалуй', 'по-видимому', 'безусловно', 'что весьма показательно'), tidy summarizing morals ('главное — не сдаваться'), balanced polished clauses, motivational-poster or moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific image, named thing, concrete action or real joke about what one can do in hard times (e.g. a particular small pleasure, everyday detail, ironic twist). Generic phrasing like 'находить радость в мелочах' or 'оставаться позитивным' with nothing distinctive is weak and caps naturalness at about 8/10. Do not credit padding or filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): No links, no meta-commentary or explanation of the task, no quoted alternatives, and the full tweet stays within 280 characters. Invented hashtags are a minor flaw, more so if not in the author's playful voice, and must never outweigh a voice or seam failure. Continuation should be in Russian."}]}, "target_words": 10, "reference": "лолировать! Жизнь слишком коротка, чтобы не смеяться. #Лолирование #Позитив ✨", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1037, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_b4d55dd6921330f2c1baa337", "id": "b4d55dd6921330f2c1baa337", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1489", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 17 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nА вы когда-нибудь пробовали хотбокс с ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends on the preposition 'с ' (with a trailing space), which needs a noun phrase in the instrumental case. Quote the joined text at the seam, e.g. 'хотбокс с [continuation start]', before deciding. Pass if the continuation begins directly with a fitting instrumental noun phrase, such as a person, a substance or a thing, and does not re-type 'с' or 'хотбокс'. It also must not open with a capitalized restart or a sentence-initial hedge like 'Ну, возможно'. A continuation that starts exactly where the prefix stops must not be marked as repeating. Penalize heavily if the case agreement is broken or the seam reads unnaturally."}, {"id": "c2", "description": "Voice and stance: The text should keep the author's casual, conspiratorial question-to-followers tone, as in 'А вы когда-нибудь пробовали...'. Russian slang, lowercase, run-ons with '..', fragments, exclamations and enthusiastic bragging or teasing all fit and are acceptable. Penalize heavily any assistant-like writing: bookish connectives or hedges ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, a tidy summarizing moral, or a moralizing or health-warning closer that breaks the author's stance. A rough but lively reply beats a correct, formal one."}, {"id": "c3", "description": "Substance: Reward a specific, concrete detail. Examples are a named companion or object to hotbox with, a vivid place such as a car, a bathroom or a balcony, a concrete consequence, or an actual punchline or absurd twist. Treat generic hyperbole ('это просто космос', 'незабываемо') and filler clauses as weak. If nothing distinctive accompanies such phrasing, cap the score at about 8/10 for naturalness."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): The output must contain only the continuation, with no links, meta-commentary, explanations of the task, or quotation marks around the answer. The full tweet, prefix plus continuation, must fit within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and should never outweigh a seam or voice failure."}]}, "target_words": 17, "reference": "друзьями на свежем воздухе? Это просто космос! Не забудьте провести небольшой конкурс на лучшее облако! #ХотбоксВулица #КуримНаПрироде", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1489, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_c81e1b8081cd6ff93e44f7b1", "id": "c81e1b8081cd6ff93e44f7b1", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:740", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nЭй друзья! Неужели это годно?! Всю ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends with 'Всю ' (a feminine accusative modifier awaiting a noun). Quote the joined text at the seam (e.g. 'Всю [first words of continuation]') before judging. The continuation must begin with a word that grammatically completes 'Всю', such as a feminine accusative noun or an adjective plus noun ('ночь', 'неделю', 'жизнь'). It must not re-type 'Всю', start with a capitalized restart, or open with a hedge. A continuation that starts directly with a fitting noun is correct and must not be counted as repetition."}, {"id": "c2", "description": "Voice and stance: the continuation keeps the author's excited, incredulous, chatty tone ('Эй друзья!', '?!') in casual Russian. Slang, exclamations, run-ons and nonstandard spelling are fine. Heavily penalize assistant-style writing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, tidy summarizing morals, or explanatory or moralizing closers. A rough but lively completion beats a correct but formal one."}, {"id": "c3", "description": "Substance: the continuation says what the author did 'всю ...' and what 'это' is, using a specific named thing, concrete action, image or real joke (e.g. a particular show, game or song binged all night). Generic hype or stock hyperbole without any distinctive detail scores no higher than about 8/10 on naturalness. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): the output contains no links, meta-commentary, or explanation of the task, and the full tweet stays within 280 characters. Invented hashtags are a minor flaw, and a playful one in the author's voice is only a slight deduction. Hygiene issues must never outweigh failures of seam or voice."}]}, "target_words": 13, "reference": "ночи смотрел это шоу, теперь у меня голова кипит от эмоций! #Годно #ШоуНочь❤️🔥", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 740, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_ced3cad5fb06614e591458a6", "id": "ced3cad5fb06614e591458a6", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:3464", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 15 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nДаша: я сама устраиваю себе терапию ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The judge must first quote the joined text at the seam (\"...устраиваю себе терапию\" + start of continuation) before deciding. The prefix ends after the noun \"терапию\" followed by a space, so a valid continuation picks up naturally from there, e.g. with a qualifier, a dash, or a description of what this 'therapy' is (\"— ...\", \"в виде...\", \"шоколадкой...\"). It must not re-type \"терапию\" or restart with a capitalized new sentence or a hedge (\"Возможно,\", \"Конечно,\"). It should read as Dasha's own quoted line continuing, or as a natural follow-up to it. A continuation that starts exactly where the prefix stops is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, self-ironic, conversational Russian register of a quoted friend's remark or a tweet about it. Lowercase, slang, fragments, '..', emoji-like enthusiasm and nonstandard spelling are fine when they fit. Penalize heavily for assistant-style writing: bookish connectives or hedges (\"по-видимому\", \"безусловно\", \"что весьма показательно\"), balanced polished clauses, and tidy morals or explanatory closers about self-care or mental health. A rough but lively line beats a correct, formal one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific image of the DIY 'therapy' or an actual joke or punchline. Examples include a named activity, food, or object, an absurd method, or a reaction from the tweet's author. Generic phrasing or stock hyperbole (\"и это помогает\", \"лучшая терапия ever\") counts as weak; cap naturalness around 8 out of 10 when nothing distinctive accompanies it. Give no credit for filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The response contains only the continuation, with no links, no meta-commentary, and no explanation of the task. The full tweet stays within 280 characters. A hashtag that sounds like the author's own playful voice is only a minor flaw. Never let a hygiene issue outweigh failures of seam, voice, or substance."}]}, "target_words": 15, "reference": "терапия ин квесчн: пьёт молоко и пляшет под бегает по полю весело кабанчик после срыва", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 3464, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_d4dc6175734cbb289c8bb1be", "id": "d4dc6175734cbb289c8bb1be", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:192", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 14 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nаааааааа вокруг так много недоразвитых которые ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text at the seam ('...недоразвитых которые' + continuation) before judging. The prefix ends on the relative pronoun 'которые', so the continuation must begin directly with that clause's predicate, typically a verb (e.g. 'не умеют...', 'думают...'). It must not repeat 'которые' or 'недоразвитых', must not open with a capitalized restart or a new sentence, and must not break grammar with 'которые'. A continuation that starts exactly at the verb is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's exasperated, ranting, informal tone, matching the opening 'аааааааа'. Lowercase, slang, mild rudeness, run-ons, '..' and emotional outbursts are fine. The author's contempt must stay intact and not be softened. Heavily penalize assistant-like writing: hedges and bookish connectives (e.g. 'возможно', 'видимо', 'безусловно', 'что показательно'), balanced polished clauses, tidy summarizing morals, or a moralizing or explanatory closer (e.g. a call for tolerance)."}, {"id": "c3", "description": "Substance: The continuation names a concrete behavior, image, or joke saying what these people actually do (a specific annoying habit, phrase, or situation). Generic insults or stock hyperbole without anything distinctive score at most about 8 on naturalness. Do not credit filler clauses or padding."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output contains no links, no meta-commentary or explanation of the task, and no invented hashtags, and the full tweet fits within 280 characters. A playful hashtag in the author's voice is only a minor flaw and must never outweigh a seam or voice failure."}]}, "target_words": 14, "reference": "до сих пор взрывают эти ебучие салюты и петарды, да идите вы нахуй, заебали", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 192, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_d5e5800fd06c8ae3582ffd04", "id": "d5e5800fd06c8ae3582ffd04", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:3834", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 16 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nмам какое будущее я тебя умоляю ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Quote the joined text where 'мам какое будущее я тебя умоляю' meets the continuation. The prefix ends on a complete exasperated phrase. The continuation should therefore read as a natural next beat, such as an elaboration, a retort to mom, or a punchline. It must not re-type 'умоляю' or other prefix words. It must not open with a capitalized formal restart or a hedge like 'Возможно' or 'Конечно же'. A continuation that simply carries on is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the teen-ish, exasperated, dismissive tone of someone arguing with their mom about 'the future'. Lowercase, slang, missing punctuation, '..' run-ons, fragments and exclamations are acceptable. Heavily penalize assistant-like writing: bookish connectives ('по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, and tidy morals or reflective closers about the meaning of the future. Also penalize any shift into a sincere, measured or optimistic stance that contradicts the author's eye-rolling position. A rough but lively line beats a correct formal one."}, {"id": "c3", "description": "Substance: Reward a concrete detail or joke that grounds the complaint. Examples include a specific exam, an absurd plan, a named everyday thing, a quoted mom phrase, or an ironic twist like 'я до завтра не доживу'. Generic hyperbole or stock phrases such as 'всё плохо' or 'какое будущее вообще' with no distinctive element are weak, and naturalness should be capped around 8. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): No links, no meta-commentary or explanation of the task, and no invented hashtags. A playful hashtag in the author's own voice is only a minor flaw. The full tweet must fit within 280 characters. These issues should never outweigh failures of seam or voice."}]}, "target_words": 16, "reference": "все что я хочу это сидеть во дворе под гитару и слушать песню про мёртвую шлюху", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 3834, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_d6ad22eb87e2565ef7d9d9cd", "id": "d6ad22eb87e2565ef7d9d9cd", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:4383", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nвообще смешно но первый вариант как ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with 'первый вариант как ' (an open comparison or relative 'как'). The judge must quote the joined text at the seam, e.g. 'первый вариант как [continuation start]'. Pass only if the continuation completes the 'как' clause grammatically, either as a comparison ('как будто…', 'как у…') or as a predicate ('как-то лучше'). It must not re-type 'как' or 'первый вариант', and must not open with a capitalized restart or a sentence-initial hedge. A continuation that starts directly where the prefix stops is correct and must not be flagged as repetition."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, lowercase, conversational Russian register of 'вообще смешно но…'. It should carry through the mixed stance: the thing is funny, but the first version was somehow better, or worse, or weirder. Slang, fragments, run-ons, missing punctuation and ')))'-style enthusiasm are fine. Heavily penalize assistant-like writing. This includes hedges and bookish connectives ('пожалуй', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, and tidy summarizing or moralizing closers. A rough but lively line beats a formal, hedged one."}, {"id": "c3", "description": "Substance: Reward a specific image, a named thing, a concrete comparison or an actual joke that says what the first version was like. Examples: a vivid simile after 'как', or a specific reference to what made it funny or better. Generic phrasing ('как-то лучше был, ну такое') or stock hyperbole with nothing distinctive caps naturalness at about 8. Filler clauses earn no credit."}, {"id": "c4", "description": "Hygiene (minor weight): Check that the output has no links, no meta-commentary or task explanation, and no invented hashtags, and that the full tweet fits in 280 characters. A playful, in-voice hashtag is only a minor flaw and must never outweigh seam or voice failures."}]}, "target_words": 13, "reference": "это выглядит для меня с экрана компа, а второй как комп сохранил мне пикчу.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 4383, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_dcc5fd4ec20bb42a45419f3f", "id": "dcc5fd4ec20bb42a45419f3f", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:3862", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 20 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nТвиттер это мои те самые скелеты ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends with 'скелеты ' (a noun followed by a space; the idiom 'скелеты в шкафу' is implied). The judge must quote the joined text at the seam, e.g. '...те самые скелеты + [first words]', before deciding. Pass: the continuation picks up naturally, for example with 'в шкафу...' or a clause that extends the thought. Fail: it re-types 'скелеты' or 'те самые', starts with a capitalized restart, or opens with a sentence-initial hedge. A continuation that simply starts where the prefix stops must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: the continuation keeps the confessional, self-ironic, colloquial Russian tone of someone calling Twitter their hidden shameful stash. Lowercase, slang, '..' run-ons, fragments, missing commas and emoji-like enthusiasm are acceptable. Heavily penalize assistant-style writing: hedges or bookish connectives ('пожалуй', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, and a tidy moral or explanatory closer (e.g. 'и это нормально, ведь каждый имеет право...')."}, {"id": "c3", "description": "Substance: reward a specific concrete detail or an actual joke about what the 'skeletons' are, such as old cringe tweets, nicknames, a reaction to a friend or mom finding the account, or a concrete action like deleting or scrolling. Generic phrasing or stock hyperbole without a distinctive detail caps naturalness at about 8. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): the output contains only the continuation, with no links, no meta-commentary or explanation of the task, and no invented hashtags. The whole tweet stays within 280 characters. A playful hashtag in the author's voice is only a minor flaw and must not outweigh seam or voice failures."}]}, "target_words": 20, "reference": "в шкафу... никто из моих друзей не подозревает с кем они на самом деле дружат.... И что я сюда выкладываю", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 3862, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_e138c662a61d58c8e3044d57", "id": "e138c662a61d58c8e3044d57", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:2157", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nМне не нужны смайлы Он к ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: Before deciding, quote the joined text at the seam ('...Он к' + continuation). The prefix ends on the preposition 'к', so the continuation must begin with its object, a noun or pronoun in the dative case (e.g. 'мне', 'тебе', a name). It must not retype 'к' or 'Он', must not restart with a capitalized new sentence, and must not open with a hedge. A continuation that starts exactly after 'к' is correct and must not be marked as repeating. The joined phrase must read as grammatical, coherent Russian and must stay consistent with 'Мне не нужны смайлы' (the emoji/attention theme)."}, {"id": "c2", "description": "Voice: The continuation keeps the author's casual, emotional, personal tweet register. Lowercase, missing punctuation, slang, run-ons with '..', and fragments are acceptable. It also keeps the author's stance, a slightly hurt or defiant attitude about smileys and someone's attention. Heavily penalize assistant-like writing: bookish connectives or hedges (e.g. 'пожалуй', 'очевидно', 'что говорит о многом'), balanced polished clauses, and tidy morals or explanatory or moralizing closers."}, {"id": "c3", "description": "Substance: Reward a concrete, specific detail about what he does or doesn't do. Examples: writing only ')' instead of words, not calling, sending stickers, coming over in person. A genuine emotional punch or joke also counts. Generic filler or stock phrasing with nothing distinctive should score no higher than about 8 out of 10 for naturalness. Padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight): Check for links, meta-commentary, translations, and explanations of the task. The full tweet must fit within 280 characters. Invented hashtags are a minor flaw, and they must never outweigh a seam or voice failure."}]}, "target_words": 13, "reference": "соседу ходит пить вино на даче а сегодня принёс от него букет роз", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 2157, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_e3dd810c00909c38613394cc", "id": "e3dd810c00909c38613394cc", "domain": "tweets", "language": "ru", "split": "validation", "group": "ru:rulexnorm:6293", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 14 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nя переключилась с о2 на Telekom. ", "rubric": {"criteria": [{"id": "c1", "description": "Seam and coherence: The prefix ends with a complete sentence ('я переключилась с о2 на Telekom. ') followed by a space, so the continuation should begin a new thought that follows naturally from switching mobile operators (e.g., reason, first impressions, complaint, comparison). The judge must quote the joined text at the seam before deciding. Penalize re-typing 'Telekom' or restating the switch as a fresh opener ('Я перешла на Telekom...'). Also penalize content unrelated to the operator switch. A continuation that simply starts a new sentence is correct and must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the author's casual, first-person feminine Russian register (lowercase, colloquial, possibly slang, fragments, '..', emoji-like enthusiasm or grumbling) and stays grammatically consistent with the feminine speaker where gender shows. Heavily penalize assistant-style writing: hedges and bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, tidy summarizing morals, or explanatory/moralizing closers. A hedged formal rewrite scores below a rough but lively one."}, {"id": "c3", "description": "Substance: Reward a concrete, specific detail tied to the switch, such as coverage in a named place, price or tariff, speed, roaming, customer service, the SIM/eSIM process, a funny mishap, or a real joke. Generic phrasing ('всё супер', 'лучшее решение в жизни') or stock hyperbole without any distinctive detail caps naturalness around 8/10. Do not credit filler clauses."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): Check that the continuation has no links, no meta-commentary or task explanation, no quoting of the prefix, is written in Russian (brand names excepted), and that the whole tweet fits within 280 characters. Invented hashtags are a minor flaw if they sound playful and in-voice, and must never outweigh a voice or seam failure."}]}, "target_words": 14, "reference": "de, и у меня даже изображения прогружаются!!! никогда так не была рада нормальной сети😭", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 6293, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_f3f19a6d8794402c663430f2", "id": "f3f19a6d8794402c663430f2", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1209", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 19 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nВот так рандомный паук заполз ко ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: The prefix ends with the preposition 'ко' followed by a space (\"Вот так рандомный паук заполз ко \"). The judge must quote the joined text at the seam, e.g. 'заполз ко мне…'. Pass if the continuation starts directly with the object of 'ко', typically 'мне' or a noun phrase such as 'мне в кровать' or 'мне на ноут', and the result is grammatical Russian. 'ко' before 'мне' is correct and must not be flagged. Fail if the continuation repeats 'ко' or earlier words, opens with a capitalized restart or a sentence-initial hedge, or produces a broken join such as 'ко в комнату'. A continuation that begins exactly where the prefix stops is not repetition."}, {"id": "c2", "description": "Voice and stance: The continuation keeps the casual, slightly ironic or startled voice of someone telling about a random spider. Lowercase, slang, '..' run-ons, fragments, mild swearing and exclamations are all acceptable. Heavily penalize assistant-style phrasing: hedges or bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что весьма показательно'), polished balanced clauses, and tidy moral or explanatory closers such as 'вот так природа напоминает о себе'."}, {"id": "c3", "description": "Substance: Reward a concrete, specific detail or an actual joke about the spider episode. Examples include where it crawled (a pillow, a keyboard, a cup of tea), what the author did (screamed, flipped a slipper over it, named it), or a funny consequence. Generic hyperbole ('я чуть не умер от страха') or vague filler with nothing distinctive caps the score around 8 of 10. Padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (weighted below c1–c3): The output must contain only the continuation, in Russian, with no links, no meta-commentary, no quotation marks wrapping the answer, and no explanation of the task. The full tweet must stay within 280 characters. A playful hashtag in the author's own voice is only a minor flaw and must never outweigh failures of seam or voice."}]}, "target_words": 19, "reference": "мне на экран... И теперь я не знаю, что делать: удалить? Убить? Или запустить его в Твиттер? #рандомный #паукижизни", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1209, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}} {"task": "tweets_ru_ffa59cdf4327634be52a4966", "id": "ffa59cdf4327634be52a4966", "domain": "tweets", "language": "ru", "split": "test", "group": "ru:rulexnorm:1847", "prompt": "Естественно допиши русскоязычный твит, сохранив позицию автора и разговорную манеру. Не повторяй начало, не добавляй ссылки, выдуманные хештеги или пояснение к заданию. Примерная длина продолжения — 13 слов; весь твит должен уложиться в 280 знаков. Верни только продолжение. Исходник сохранён с токенизацией корпуса; необычная орфография сама по себе не ошибка качества.\n\nНачало:\nмальчик мой да ты хоть знаешь ", "rubric": {"criteria": [{"id": "c1", "description": "Seam: the prefix ends with \"мальчик мой да ты хоть знаешь \" — mid-clause, after the verb 'знаешь', which needs an object or a clause (e.g. 'что', 'как', 'сколько', a noun). The judge must quote the joined text at the seam, e.g. '...ты хоть знаешь [continuation start]', before deciding. Pass if the continuation directly supplies what is known, without retyping 'знаешь' or 'ты хоть'. It must not open with a capitalized restart, a new sentence, or a hedge. A continuation starting exactly at that point must not be marked as repeating."}, {"id": "c2", "description": "Voice and stance: the text keeps the affectionate, mock-exasperated or teasing address ('мальчик мой') in colloquial Russian. Lowercase, slang, run-ons, '..', exclamations and nonstandard spelling are acceptable. Heavily penalize assistant-like writing: hedges and bookish connectives ('возможно', 'по-видимому', 'безусловно', 'что показательно'), balanced polished clauses, a tidy moral, or an explanatory or moralizing closer. A rough but lively line beats a formal one."}, {"id": "c3", "description": "Substance: the continuation names a specific thing the boy supposedly doesn't know or appreciate. Examples are a concrete fact, an action, a named object or person, or an actual punchline. Generic hyperbole ('как я тебя люблю', 'как всё сложно') with nothing distinctive caps quality around 8/10. Filler or padding clauses earn no credit."}, {"id": "c4", "description": "Hygiene (lower weight than c1–c3): the output contains no links, no meta-commentary, and no explanation of the task. The whole tweet fits in 280 characters. A hashtag in the author's playful voice is only a minor flaw and must not outweigh voice or seam failures."}]}, "target_words": 13, "reference": "сколько раз до тебя мне уже отказали дебатеры спикеры и площадки для мероприятий.", "license": "CC-BY-SA-2.5", "source": {"url": "https://raw.githubusercontent.com/SmartEngines/RuLexNorm/59a8ebc3d314fafe25f918f50e5048f8ac00bed8/data/RuLexNorm.norm", "revision": "59a8ebc3d314fafe25f918f50e5048f8ac00bed8", "license": "CC-BY-SA-2.5 (repository README)", "attribution": "Irina Koliaskina and Dmitry Sholomov / Smart Engines / RuLexNorm", "entry_index": 1847, "source_language": "ru", "normalized_column_used": false, "text_transform": "Original left-column tokens joined by single spaces; original whitespace unavailable.", "thread_or_author_metadata_available": false, "quality": "Real historical Twitter utterance; not expert writing or a confirmed standalone tweet."}}