Datasetpublic

marin-community/open-thoughts-4-30k-math-qwen3-4b-annotated-32768-tokens-n8-rejection-sampling-soft-match

N8 Rejection Sampling (Soft Match) Overview This dataset was created via rejection sampling from the Qwen3-4B response dataset using Qwen3-32B answers as ground truth. Source dataset (Qwen3-4B, 8 responses per prompt): marin-community/open-thoughts-4-30k-math-qwen3-4b-annotated-32768-tokens-n8-reformatted Verifier dataset (Qwen3-32B, 1 response per prompt): marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens Creator: The Marin Project… See the full description on the dataset page: https://huggingface.co/datasets/marin-community/open-thoughts-4-30k-math-qwen3-4b-annotated-32768-tokens-n8-rejection-sampling-soft-match.

sourceHugging Faceupdated 7mo agoView on Hugging Face

0likes102downloads

Dataset Card

N8 Rejection Sampling (Soft Match)

Overview

This dataset was created via rejection sampling from the Qwen3-4B response dataset using Qwen3-32B answers as ground truth.

—Source dataset (Qwen3-4B, 8 responses per prompt): marin-community/open-thoughts-4-30k-math-qwen3-4b-annotated-32768-tokens-n8-reformatted
—Verifier dataset (Qwen3-32B, 1 response per prompt): marin-community/open-thoughts-4-30k-math-qwen3-32b-annotated-32768-tokens
—Creator: The Marin Project

How This Dataset Was Created

For each of the 29,963 unique prompts in the source dataset (which has 8 responses per prompt), we checked whether any of the 8 Qwen3-4B final answers matched the Qwen3-32B final answer for the same prompt. If at least one match was found, we kept only the matching samples and discarded the rest. If no matches were found, we kept all 8 samples (rather than discarding them). If the Qwen3-32B answer was "N/A", we kept all 8 samples.

This is a softer filtering: verified responses are preferred, but unverifiable prompts are retained in full.

Prompts were matched between the two datasets using the ms_id column (unique per prompt). Final answers were compared via exact string match on the final_answer column.

Rejection Sampling Statistics

Metric	Count
Total unique prompts	29,963
Prompts with at least one Qwen3-4B answer matching Qwen3-32B	17,932
Prompts with zero matches between Qwen3-4B and Qwen3-32B	8,653
Prompts where Qwen3-32B answer was "N/A"	3,378
Prompts where all 8 Qwen3-4B answers were "N/A"	1,295
Final dataset size	187,194 rows

Schema/Columns

Column	Type	Description
`row_id`	int64	Unique identifier
`instruction_seed`	string	Original instruction/prompt
`_source`	string	Data source identifier
`gpt41_mini_response`	string	Reference response
`__original_row_idx`	int64	Index from original dataset
`length`	int64	Response length
`ms_id`	int64	Mapping ID (unique per prompt)
`generated_text`	string	Model-generated response with reasoning trace
`final_answer`	string	Extracted final answer from \\boxed{}
`complete_responses_count`	int64	Number of complete responses for this prompt

Access

python

from datasets import load_dataset

dataset = load_dataset("marin-community/open-thoughts-4-30k-math-qwen3-4b-annotated-32768-tokens-n8-rejection-sampling-soft-match")
print(dataset["train"][0])

Use Cases

—Training models on mathematical reasoning with verified responses
—Fine-tuning for chain-of-thought reasoning
—Studying the effect of rejection sampling on training data quality
—Comparing strict vs. soft filtering strategies for synthetic data