README.md

metadata

license: apache-2.0
language:
  - en
task_categories:
  - text-generation
configs:
  - config_name: questions
    default: true
    data_files:
      - split: train
        path: data/questions/train.parquet
  - config_name: with_answers
    data_files:
      - split: train
        path: data/with_answers/train.parquet

OpenThoughts3 processed questions

This dataset extracts the first conversations[0].value from each row of OpenThoughts3-1.2M, deduplicates questions by their original text, and filters lexical overlaps with the benchmarks below. The original difficulty, source, and domain fields are retained from the first occurrence. Each conversations array contains only the original question dictionary; generated responses have been removed.

Configuration	Columns	Rows
questions	difficulty, source, domain, conversations	65,064
with_answers	the four columns above plus answer, answer_method, answer_confidence, response_count, answer_support	65,064

Domain counts in both configurations: math 53,100, code 5,714, science 6,250. These are measured counts after filtering; the code count is substantially below 16,000.

Processing

• Source: all 120 train shards (1,200,000 rows) at revision 61bcf9d4eb38b30295efc2021227a63cc5bb34c8.
• Exact duplicate questions removed: 1,134,911 rows. Benchmark overlapping unique questions removed: 25. Final unique questions: 65,064.
• Benchmarks: MATH-500 (500), AIME 2025 (30), AIME 2026 (30), HMMT February 2025 (30), HMMT February 2026 (33), IMO-AnswerBench v2 (400), LiveCodeBench release_v6 (1,055 cumulative problems from test.jsonl through test6.jsonl), and OJBench prompts (464 English/Chinese prompts).
• Contamination matching uses Unicode NFKC, case folding, alphanumeric normalization, exact normalized text, and word-anchor candidate retrieval followed by containment or fuzzy similarity. All 25 removed questions matched MATH-500. This is a lexical screen and cannot certify the absence of semantic paraphrases.

Extracted answers

The with_answers configuration mines only completed final sections after </think> from the original generated responses. For math, it uses final \boxed{...} or an explicit final-answer label and selects the most frequent extracted answer across responses; ties become null. For code, it takes a final fenced code block. For science, it retains a completed final response as the answer text. Questions with no explicit completed answer retain null.

Domain	Extracted answer	null
Math	29,840	23,260
Code	4,464	1,250
Science	6,066	184

answer_confidence describes extraction evidence, not correctness: high means at least two math responses agree on the selected text, medium means a single explicit math answer or a completed code/science answer was extracted, and none means answer is null. response_count counts source assistant responses for the question; answer_support counts responses supporting a selected math answer (and is 1 for extracted code/science answers). Generated answers have not been independently checked against ground truth or executed as code. Treat them as answer candidates.

Load

from datasets import load_dataset

questions = load_dataset("ChuGyouk/OpenThoughts3-processed", "questions", split="train")
with_answers = load_dataset("ChuGyouk/OpenThoughts3-processed", "with_answers", split="train")

See processing_report.json for counts and process.py in this repository for the extraction rules.