id stringlengths 16 20 | domain stringclasses 1
value | model stringclasses 1
value | problem_id stringlengths 4 8 | sample int64 0 3 | question stringlengths 755 4.34k | status stringclasses 2
values | final_answer stringlengths 0 129k | messages listlengths 3 3 | completion_tokens int64 766 81.9k | generation_config dict | benchmark stringclasses 1
value | platform stringclasses 3
values | difficulty stringclasses 3
values | contest_date timestamp[s]date 2023-05-13 00:00:00 2025-04-06 00:00:00 | code stringlengths 0 26.3k | verification dict |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
code-4b-1899_C__s3 | code | Qwen/Qwen3.5-4B | 1899_C | 3 | "### Question:\nA subarray is a continuous part of array.\n\nYarik recently found an array $a$ of $n(...TRUNCATED) | reply_cut | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 81,920 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | codeforces | hard | 2023-10-17T00:00:00 | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"no_code","tests(...TRUNCATED) | ||
code-4b-2754__s3 | code | Qwen/Qwen3.5-4B | 2754 | 3 | "### Question:\nYou are given a 0-indexed integer array nums representing the score of students in a(...TRUNCATED) | done | "```python\nimport math\nfrom typing import List\n\nclass Solution:\n def maxStrength(self, nums:(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 16,433 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | medium | 2023-05-27T00:00:00 | "import math\nfrom typing import List\n\nclass Solution:\n def maxStrength(self, nums: List[int])(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
code-4b-2784__s0 | code | Qwen/Qwen3.5-4B | 2784 | 0 | "### Question:\nYou are given a 0-indexed integer array nums representing the strength of some heroe(...TRUNCATED) | done | "```python\nfrom typing import List\n\nclass Solution:\n def sumOfPower(self, nums: List[int]) ->(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 18,179 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | hard | 2023-05-13T00:00:00 | "from typing import List\n\nclass Solution:\n def sumOfPower(self, nums: List[int]) -> int:\n (...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
code-4b-2786__s1 | code | Qwen/Qwen3.5-4B | 2786 | 1 | "### Question:\nYou are given a 0-indexed string s that consists of digits from 0 to 9.\nA string t (...TRUNCATED) | done | "```python\nclass Solution:\n def longestSemiRepetitiveSubstring(self, s: str) -> int:\n n(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 19,031 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | medium | 2023-06-10T00:00:00 | "class Solution:\n def longestSemiRepetitiveSubstring(self, s: str) -> int:\n n = len(s)\n(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
code-4b-2817__s1 | code | Qwen/Qwen3.5-4B | 2817 | 1 | "### Question:\nYou are given a 0-indexed binary string s of length n on which you can apply two typ(...TRUNCATED) | done | "```python\nclass Solution:\n def minimumCost(self, s: str) -> int:\n n = len(s)\n (...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 894 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | medium | 2023-05-28T00:00:00 | "class Solution:\n def minimumCost(self, s: str) -> int:\n n = len(s)\n # Calculate(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
code-4b-2817__s2 | code | Qwen/Qwen3.5-4B | 2817 | 2 | "### Question:\nYou are given a 0-indexed binary string s of length n on which you can apply two typ(...TRUNCATED) | done | "```python\nclass Solution:\n def minimumCost(self, s: str) -> int:\n n = len(s)\n (...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 1,140 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | medium | 2023-05-28T00:00:00 | "class Solution:\n def minimumCost(self, s: str) -> int:\n n = len(s)\n cost = 0\n (...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
code-4b-2839__s0 | code | Qwen/Qwen3.5-4B | 2839 | 0 | "### Question:\nYou are given two 0-indexed integer arrays nums1 and nums2, each of length n, and a (...TRUNCATED) | done | "```python\nimport sys\nfrom typing import List\nfrom bisect import bisect_left\n\nclass Solution:\n(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 21,664 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | hard | 2023-06-11T00:00:00 | "import sys\nfrom typing import List\nfrom bisect import bisect_left\n\nclass Solution:\n def max(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"time_limit","te(...TRUNCATED) |
code-4b-2849__s3 | code | Qwen/Qwen3.5-4B | 2849 | 3 | "### Question:\nThe imbalance number of a 0-indexed integer array arr of length n is defined as the (...TRUNCATED) | done | "```python\nfrom typing import List\n\nclass Solution:\n def sumImbalanceNumbers(self, nums: List(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 34,405 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | hard | 2023-07-02T00:00:00 | "from typing import List\n\nclass Solution:\n def sumImbalanceNumbers(self, nums: List[int]) -> i(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"runtime_error",(...TRUNCATED) |
code-4b-2850__s3 | code | Qwen/Qwen3.5-4B | 2850 | 3 | "### Question:\nYou are given three integers x, y, and z.\nYou have x strings equal to \"AA\", y str(...TRUNCATED) | reply_cut | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 81,920 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | medium | 2023-06-24T00:00:00 | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"no_code","tests(...TRUNCATED) | ||
code-4b-2884__s2 | code | Qwen/Qwen3.5-4B | 2884 | 2 | "### Question:\nYou are given a string word and an array of strings forbidden.\nA string is called v(...TRUNCATED) | done | "```python\nimport collections\nfrom typing import List\n\nclass Solution:\n def longestValidSubs(...TRUNCATED) | [{"role":"system","content":"You are an expert Python programmer. You will be given a question (prob(...TRUNCATED) | 46,546 | {"sampling":{"temperature":1.0,"top_p":0.95,"top_k":20,"min_p":0.0,"presence_penalty":1.5},"limits":(...TRUNCATED) | LiveCodeBench release_v6 | leetcode | hard | 2023-07-16T00:00:00 | "import collections\nfrom typing import List\n\nclass Solution:\n def longestValidSubstring(self,(...TRUNCATED) | {"grader":"LiveCodeBench hidden tests, regraded twice at low parallelism","verdict":"wrong_answer","(...TRUNCATED) |
End of preview. Expand in Data Studio
qwen3.5-inferences: verified wrong first attempts (Qwen3.5-4B / 9B; code, math, search)
3,000 failed first attempts for learning-from-failure / retry experiments: 500 per domain × model, one file each. Every sample is the model's complete first attempt (reasoning, final reply and, for search, every tool call and result). Every sample has been re-verified as wrong.
| File | Benchmark | Samples | Distinct problems |
|---|---|---|---|
| code-4b.jsonl | LiveCodeBench release_v6 (May 2023 – Apr 2025, 1,055 problems) | 500 | 434 |
| code-9b.jsonl | same | 500 | 309 |
| math-4b.jsonl | all MathArena 2025–2026 final-answer sets + arXivMath (675 problems) | 500 | 474 |
| math-9b.jsonl | same | 500 | 440 |
| search-4b.jsonl | BrowseComp-Plus, all 830 queries, BM25 retriever | 500 | 500 |
| search-9b.jsonl | same | 500 | 500 |
How the samples were made
- Each problem was sampled 4 times (temperature 1.0, top-p 0.95, top-k 20, presence penalty 1.5; context 128K;
generation_configin each row has the exact settings). The released samples are failed ones from that pool. - Selection: up to 500 failures per file, spread over as many problems as possible (round robin over problems, seed 0).
How "wrong" was verified
- Code: every failure was regraded twice on the LiveCodeBench hidden tests, one program at a time. A sample is
included only if it fails both times. (The original grading ran 128 programs in parallel and wrongly failed many
correct programs, which is why this was needed: 515 of 1,640 4B "failures" and 597 of 1,392 9B "failures" were
actually correct and are excluded.)
verification.verdictis one of no_code, wrong_answer, time_limit, runtime_error. - Math: regraded with math-verify against the gold answer. Answers that a Qwen3.5-9B equivalence check flagged as possibly equivalent (4 of 797) were checked by hand; all 4 are wrong. Samples with no final answer (cut off or unparsable) are included as wrong.
- Search: judged by Qwen3.5-27B, shown only the text inside the last
<answer>tag, against the gold answer. Samples with no<answer>tag (e.g. the turn limit was reached) are included as wrong.
Fields
id, domain, model, problem_id, sample, question (the user prompt), status (done / reply_cut / max_turns /
context_full), final_answer (the final reply), messages (all turns: role, content, reasoning_content, tool_calls,
finish_reason, token counts), completion_tokens, generation_config, verification; plus per domain:
- code:
benchmark,platform,difficulty,contest_date,code(the extracted program). Hidden tests are not included; get them from LiveCodeBench byproblem_id(= question_id). - math:
benchmark,set,gold_answer,extracted_answer. - search:
benchmark,gold_answer.
Notes
- The search files contain BrowseComp-Plus questions and gold answers, and all files contain benchmark problems. This is evaluation data: do not train on it.
- Code first-try accuracy after regrading: about 73% (4B) and 81% (9B), so code failures are mostly near-misses on problems the model often solves.
- Downloads last month
- -