Hugging Face's logo

Datasets:
JingweiNi
/
ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513

Tasks:
Text Generation
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
< 1K
Tags:
reasoning
code
codeforces
opencode-reasoning
k2-think
qwen
Libraries:
Datasets
Dask
Polars
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer
Auto-converted to Parquet Duplicate
question
stringlengths
514
5.71k
answer
stringclasses
1 value
question_id
stringlengths
32
32
ocr2_id
stringlengths
32
32
language_split
stringclasses
2 values
source_dataset
stringclasses
1 value
source_split
stringclasses
2 values
source_platform
stringclasses
1 value
difficulty
stringdate
1900-01-01 00:00:00
3500-01-01 00:00:00
judgement
stringclasses
1 value
pass_rate
stringclasses
72 values
solution
stringclasses
1 value
qwq_critique
stringclasses
1 value
source_index
stringlengths
1
5
input_ids
listlengths
4.44k
187k
reply
stringlengths
14.5k
605k
sample_idx
int64
0
0
original_index
int64
0
499
claims
listlengths
189
7.8k
verified
listlengths
189
7.8k
"Arthur owns a ski resort on a mountain. There are n landing spots on the mountain numbered from 1 t(...TRUNCATED)
aaf7b7ea051db7ebebecc77c63408b08
829df77971430ea99e518852a698e82c
python
code_contests
train
codeforces
2500
right
0.0
7729
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to solve the problem described: Given a DAG (one-directional tracks all descending index wi(...TRUNCATED)
0
0
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23],"claim_text":"(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"You are given an array a_1, a_2, ..., a_n. All a_i are pairwise distinct.\n\nLet's define function (...TRUNCATED)
5243682906dc6806d5dbc0884dd153e9
397c784a9e13c36d300def6acfd896bc
python
code_contests
train
codeforces
2300
right
1.0
9281
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need sum over all subsegments (subarrays) of f(l,r), where f(l,r) is the sum of sorted subarray (...TRUNCATED)
0
1
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"You are given an integer n from 1 to 10^{18} without leading zeroes.\n\nIn one move you can swap an(...TRUNCATED)
d0d6b21da715f9a948cf95c6245415da
c35088ddf60ab7295b28104312783d28
python
code_contests
train
codeforces
2100
right
0.9
5379
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We are given an integer n (as a decimal string) up to length 19 digits (since n ≤ 10^18). Actuall(...TRUNCATED)
0
2
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23],"claim_text":"(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"The crowdedness of the discotheque would never stop our friends from having fun, but a bit more spa(...TRUNCATED)
a791bf16c4674794d022cd2b5c8d4c66
4343056fd4e987758dd78f55910d6997
python
code_contests
train
codeforces
2000
right
0.46
1521
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We have an interesting problem: we have n dancers with circular movement ranges (disks). They can b(...TRUNCATED)
0
3
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23],"claim_text":"(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"Boy Dima gave Julian a birthday present — set B consisting of positive integers. However, he didn(...TRUNCATED)
0823ef3ce73f71c507e9e58d06f195db
f2351b20573037dc38a94e5febc2e689
python
code_contests
train
codeforces
1900
right
0.96
10846
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to solve this problem. Let's restate:\n\nGiven a set B of positive integers (size n up to 2(...TRUNCATED)
0
4
[{"aligned_token_ids":[0,1,2,3,4,5,6,7],"claim_text":"We need to solve this problem.","sentence":"We(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"On the board, Bob wrote n positive integers in [base](https://en.wikipedia.org/wiki/Positional_nota(...TRUNCATED)
e87cbb097a65b8b734d98ee3b2eced9e
579ba8bdf3f9b45e42ab6cd778c588b4
python
code_contests
valid
codeforces
2000
right
0.08
95
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to understand the problem fully. Bob writes n positive integers (in base 10) whose sum equa(...TRUNCATED)
0
5
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9],"claim_text":"We need to understand the problem fully.",(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"You are given an array a_1, a_2, ..., a_n. You can perform the following operation any number of ti(...TRUNCATED)
0d05ce91e5fbfbdd812c17b48c17f24f
11dc394188b1a67f4f76d68c5075d89d
python
code_contests
train
codeforces
2100
right
1.0
4917
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to solve the problem: given an array a[1..n] with elements up to 1000, n up to 500. Operati(...TRUNCATED)
0
6
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13],"claim_text":"We need to solve the problem: (...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"The government of Berland decided to improve network coverage in his country. Berland has a unique (...TRUNCATED)
e21b0d26bb9ffbfc74dc52bfaf8901a6
1374d24b05b2a099bb48093445a6350e
python
code_contests
train
codeforces
2400
right
1.0
9604
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to solve the described problem and produce the answer in the given format.\n\nWe need to an(...TRUNCATED)
0
7
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25],"claim_t(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"Polycarp is making a quest for his friends. He has already made n tasks, for each task the boy eval(...TRUNCATED)
aa8c320b16891d6d9570562cfadec8b5
d8e4318417351a0abb1be7aa203323c1
python
code_contests
train
codeforces
2100
right
1.0
11919
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to interpret the problem precisely and find an algorithm.\n\nWe have n tasks, each with int(...TRUNCATED)
0
8
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15],"claim_text":"We need to interpret the(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
"Dima took up the biology of bacteria, as a result of his experiments, he invented k types of bacter(...TRUNCATED)
08cfc541b501482be83ba54e54301dbb
f4e9657d1f3120f83e477841dbcc8f94
python
code_contests
train
codeforces
2000
right
0.8125
8788
[0,250018,5671,256,54705,1222,4778,4920,38,2211,31417,2204,8506,1245,1919,44766,6719,44766,108490,52(...TRUNCATED)
"We need to understand the problem statement precisely. There's a summary: Dima invented k types of (...TRUNCATED)
0
9
[{"aligned_token_ids":[0,1,2,3,4,5,6,7,8,9,10,11,12,13],"claim_text":"We need to understand the prob(...TRUNCATED)
[null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null,null(...TRUNCATED)
End of preview. Expand in Data Studio

Qwen3.5 FP8 Annotations for 10K K2-Think OCR2 Coding Steps, Sample500

This dataset contains Qwen3.5 FP8 step-level correctness annotations for K2-Think reasoning traces on a diverse 500-problem hard Codeforces subset reconstructed from OpenCodeReasoning-2/code_contests.

Summary

  • Source trace dataset: opencode_reasoning2_hard_codeforces1900_pr10_reconq_k2_thinking_extracted_sample500_seed20260514
  • Source rows: 500 coding problem traces
  • Candidate step rule: claim with non-empty aligned_token_ids
  • Candidate aligned steps: 1590686
  • Manifest-selected annotated steps: 10000
  • Sampling seed: 20260513
  • K2 problem-sampling seed: 20260514
  • Relative-position deciles: exactly 1,000 selected steps in each decile 0 through 9
  • Annotator: Qwen/Qwen3.5-122B-A10B-FP8
  • Served model name: Qwen/Qwen3.5-122B-A10B
  • Final valid selected labels: 10000
  • Final invalid selected labels: 0
  • Qwen-labelled K2 error labels: 1536
  • Qwen-labelled K2 error rate among selected steps: 0.1536

The dataset has 500 rows, one per coding problem trace. The verified column is a list aligned with claims. For the 10,000 manifest-selected steps, verified contains a numeric label:

  • 0.0: Qwen judged the step correct
  • 1.0: Qwen judged the step incorrect

For non-selected claim steps, verified is NaN. This is intentional; only the manifest-selected steps should be treated as annotated.

Source Universe

The K2-Think traces were generated from a deterministic random sample of 500 Codeforces coding problems using seed 20260514. The final extracted universe contains 1590686 usable aligned candidate steps across 500 traces.

Filtering/sampling provenance:

  • OCR2 language splits: python,cpp, deduplicated by question_id
  • Source dataset: code_contests
  • Source platform: codeforces
  • Minimum rating: 1900
  • Maximum pass rate: 1.0
  • Judgement filter: right
  • Problem statements were reconstructed from deepmind/code_contests; the placeholder OCR2 question field was not used for final trace generation.

Manifest Sampling

The 10,000 annotated steps were sampled with a deterministic manifest:

  • Candidate id is a stable hash of dataset version, question_id, original_index, step_idx, and claim_text.
  • Position bucket is one of 10 relative-position deciles using (step_idx + 0.5) / num_valid_claims_in_trace.
  • Exactly 1,000 unused steps were selected per decile.
  • Prior pilot annotations and prior manifests were excluded before sampling.

The manifest and run summaries are included under metadata/.

Clariden Run

  • Clariden account: a0142
  • Environment: qwen35_vllm
  • Qwen vLLM args: --tensor-parallel-size 4 --gpu-memory-utilization 0.90 --trust-remote-code --reasoning-parser qwen3
  • Confirmed server properties: max_seq_len=262144, enforce_eager=False
  • Main annotation array job: 2259695
  • Initial merge job: 2259696
  • Recovery annotation job for 38 initially invalid selected labels: 2262035
  • Recovery merge job: 2262036

The first merge produced 9,962 valid selected labels and 38 selected NaN labels because some Qwen responses followed problem-embedded output instructions instead of returning the verifier digit. Those 38 entries were retried with a hardened verifier prompt and strict 0/1 retry path, then re-merged. The final report validates 10,000 selected labels and 0 invalid selected labels.

Columns

The dataset preserves the source coding-problem and trace fields, including:

  • question
  • answer
  • question_id
  • ocr2_id
  • source_dataset
  • difficulty
  • solution
  • reply
  • claims
  • verified

Each claims entry contains the claim text and aligned token ids for a reasoning step.

Intended Use

This dataset is intended for experiments on step-level verification, uncertainty estimation, and process supervision for coding-reasoning traces.
Downloads last month
38
Number of rows:
500
Total file size:
368 MB