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.
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
- A response is complete if it contains a
</think>token (end of reasoning trace) followed by a non-empty final answer enclosed in\boxed{...}. - For each prompt (group of 8 responses), the final answers of all complete responses are compared via majority vote.
- Kept: Responses whose final answer matches the majority vote winner (must have >= 2 votes).
- Discarded entirely (all 8 responses removed):
- Prompts with no complete responses
- Prompts with only 1 complete response
- Prompts where no answer received >= 2 votes (no majority winner)
- Discarded individually: Incomplete responses, and complete responses that disagree with the majority.
Filtering Statistics
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
Row Ordering
- Rows are ordered so that
ms_idordering matches the reference dataset marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens. - All responses for a given prompt appear contiguously.
Related Datasets
- open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8-reformatted — Pre-filtering version with all 239,704 rows
- open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens-n8 — Original dataset with 8 response columns per row
- open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens — Single-response version (n=1) with max_tokens = 32768
