OpenR1-Math-10k-Raw

# OpenR1-Math-10k-Raw ## What this dataset is This dataset is a deterministic 10,000-example raw subset derived from the `all` configuration of [open-r1/OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k). It is intended as a shared upstream artifact for later: - SFT dataset derivation - GRPO dataset derivation - evaluation and audit tracing The goal is to keep one common raw reasoning dataset and only derive training-format-specific views later, instead of mixing separate upstream sources for SFT and GRPO. ## How it was created The subset was created from the `train` split of the `all` config with a deterministic seeded selection policy. Source dataset facts recorded during extraction: - source dataset: `open-r1/OpenR1-Math-220k` - source config: `all` - source split: `train` - source rows seen: `225,129` - target subset size: `10,000` - seed: `42` Filtering rules applied before selection: - keep rows with non-empty `problem`, `solution`, and `answer` - require `correctness_count >= 1` - require at least one `true` value in `is_reasoning_complete` - exclude `question_type == "proof"` After filtering, deterministic sampling was performed using a stable hash of `seed` plus `uuid` or `problem`. ## Fields Each row preserves the raw fields needed for later downstream conversion: - `problem` - `solution` - `answer` - `problem_type` - `question_type` - `source` - `uuid` - `is_reasoning_complete` - `generations` - `correctness_math_verify` - `correctness_llama` - `finish_reasons` - `correctness_count` - `messages` ## Intended downstream usage ### SFT Typical derivation: - input: `problem` - target: `solution` ### GRPO Typical derivation: - prompt: `problem` - reward reference: `answer` This repo intentionally keeps those later conversions separate from the raw subset. ## Important limitations - This is a derived subset, not the full upstream OpenR1-Math corpus. - It is a reasoning-focused raw dataset, not a ready-made SFT or GRPO dataset. - The filtering policy may bias the subset toward more verifiable and more complete reasoning traces. - The subset should be treated as a project-specific research artifact, not a replacement for the original upstream dataset. ## Provenance - Upstream dataset: [open-r1/OpenR1-Math-220k](https://huggingface.co/datasets/open-r1/OpenR1-Math-220k) - Project repo: `do-we-regret-lora`