CoolFace
Datasetpublic

marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted-selfconsistency

Dataset Card for Open-Thoughts-4-30K-Math-Qwen3-32B-Annotated-32768-Tokens-N8-Reformatted-SelfConsistency Overview This dataset is a self-consistency filtered version of marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted. For each prompt, 8 responses were generated by Qwen3-32B with different random seeds. A majority vote was taken over the final answers (extracted from \boxed{...}) to determine the most popular answer, and… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted-selfconsistency.

sourceHugging Faceupdated 8mo agoView on Hugging Face
3likes199downloads
Dataset Card

Dataset Card for Open-Thoughts-4-30K-Math-Qwen3-32B-Annotated-32768-Tokens-N8-Reformatted-SelfConsistency

Overview

This dataset is a self-consistency filtered version of marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted. For each prompt, 8 responses were generated by Qwen3-32B with different random seeds. A majority vote was taken over the final answers (extracted from \boxed{...}) to determine the most popular answer, and only responses agreeing with that majority answer were kept.

Filtering Criteria

  1. 1.A response is complete if it contains a </think> token (end of reasoning trace) followed by a non-empty final answer enclosed in \boxed{...}.
  2. 2.For each prompt (group of 8 responses), the final answers of all complete responses are compared via majority vote.
  3. 3.Kept: Responses whose final answer matches the majority vote winner (must have >= 2 votes).
  4. 4.Discarded entirely (all 8 responses removed):
  5. 5.Prompts with no complete responses
  6. 6.Prompts with only 1 complete response
  7. 7.Prompts where no answer received >= 2 votes (no majority winner)
  8. 8.Discarded individually: Incomplete responses, and complete responses that disagree with the majority.

Filtering Statistics

MetricCount
Original prompt groups29,963
Kept prompt groups27,760 (92.6%)
Discarded (no complete responses)367
Discarded (only 1 complete response)372
Discarded (no majority winner with >= 2 votes)1,464
Total kept rows154,407
Discarded rows (incomplete)27,483
Discarded rows (wrong answer)57,814

Generation Details

  • —Model: Qwen/Qwen3-32B
  • —Temperature: 0.8
  • —Max Output Tokens: 32768
  • —Responses per prompt: 8 (different random seeds)

Dataset Statistics

  • —Number of Samples: 154,407
  • —Number of Unique Prompts: 27,760
  • —Average responses kept per prompt: ~5.6

Dataset Structure

ColumnDescription
row_idA unique row identifier (0 to 154,406)
instruction_seedThe original math problem/question text without chat formatting
_sourceThe origin dataset (e.g., ai2-adapt-dev/openmath-2-math); tracks data provenance
gpt41_mini_responseA reference solution generated by GPT-4.1 Mini
__original_row_idxThe row index from the original source dataset before filtering/processing
lengthThe token count of the response in gpt41_mini_response
ms_idA unique sample identifier (shared across responses for the same prompt)
generated_textA response including chain-of-thought with <think> tags, generated by Qwen3-32B
final_answerThe majority-vote winning answer extracted from \boxed{...}
majority_vote_agree_countNumber of complete responses that agreed with the majority answer
complete_responses_countTotal number of complete responses (out of 8) for this prompt

Row Ordering

Related Datasets