This is a filtered and metadata enriched version of open-thoughts/OpenThoughts-114k.

While the original dataset is a valuable resource containing DeepSeek-R1 outputs, it has very little metadata (only 2 fields: system and conversations). It does not contain, for instance, the original solution label, which means that we can not verify the model answers.

What we did

• filtered the dataset for math content (math questions were prefixed by "Return your final response within \boxed{}." -- see here)
• found the original questions in the AI-MO/NuminaMath-CoT and mapped them back to each generation
• verified model generations using our Math-Verify library
• added a metadata field with the token count of each DeepSeek-R1 completion

Data structure

• source: original source from Numina-Math
• problem: problem statement, from Numina-Math
• solution: original solution/gold label, from Numina-Math
• messages: message turns for finetuning on the correct solutions, from Numina-Math
• system: system prompt sent to DeepSeek-R1, from OpenThoughts
• conversations: message turns from the DeepSeek-R1 generation. The last turn is the model output, from OpenThoughts
• generated_token_count: number of tokens (counted using the DeepSeek-R1 tokenizer) of the model output.
• correct: label indicating if the DeepSeek-R1 generated solution matches the ground truth solution. Checked with Math-Verify library

Some statistics

• The original OpenThoughts-114k dataset has 89120/113957 (78%) math rows
• Of those, 56730/89120 (63%) have correct answers, as checked by Math-Verify
• There is a single generation per question
• Token count distribution: mean=6366.67, std_dev=4662.88 tokens