Competitive-Programming-python-blend

# Dataset Card for Competitive-Programming-python-blend ## Summary `Competitive-Programming-python-blend` is a mixed supervised fine-tuning dataset centered on competitive programming, code reasoning, and instruction-style problem solving. The blend is Python-first, but it also keeps a small amount of C++, agentless SWE, and reasoning-oriented chat supervision to broaden training coverage. The current release is published as a single HF-friendly JSONL file, `clean.jsonl`. ## Blend Composition The blend follows the source proportions below. | Source | Role in the blend | Share | | --- | --- | ---: | | [nohurry/Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered) | Reasoning-heavy synthetic SFT data | 5.83% | | [Jackrong/Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x) | Distilled reasoning and instruction-following data | 1.58% | | [nvidia/Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2), `competitive_coding_python` | Primary Python competitive-programming supervision | 87.54% | | [nvidia/Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2), `competitive_coding_cpp` | Small cross-language competitive-programming supplement | 2.50% | | [nvidia/Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2), `agentless` | Lightweight agentless SWE-style supervision | 0.05% | | [nvidia/Nemotron-SFT-Instruction-Following-Chat-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Instruction-Following-Chat-v2), `reasoning_on` | Small reasoning-oriented chat supplement | 2.50% | Percentages are computed from the blend recipe and sum to 100%. ## Data Format Each line in `clean.jsonl` is one JSON object with a `messages` field. The current release is stored in a format that can be loaded directly with `datasets.load_dataset("json", ...)`: ```json { "id": "e3f7b0d4f8fbb2f33771b2d8f0cbecab6d5e3f1b85f58fca4d3fbf5ce7d8f98b", "messages": [ {"role": "user", "content": "prompt"}, {"role": "assistant", "content": "<think>..."} ] } ``` All upstream records were normalized into this unified schema. Source-specific fields were flattened into text turns, speaker names were standardized into `user` / `assistant` style roles, and each sample carries a content-derived SHA-256 `id` string. ### Source-by-source notes - [nohurry/Opus-4.6-Reasoning-3000x-filtered](https://huggingface.co/datasets/nohurry/Opus-4.6-Reasoning-3000x-filtered): upstream card declares `apache-2.0`. No dedicated citation block is provided on the card. - [Jackrong/Qwen3.5-reasoning-700x](https://huggingface.co/datasets/Jackrong/Qwen3.5-reasoning-700x): upstream card declares `apache-2.0`. The card also notes that usage should comply with the Qwen open-source license agreement and Alibaba Cloud DashScope terms. No dedicated citation block is provided on the card. - [nvidia/Nemotron-SFT-Competitive-Programming-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Competitive-Programming-v2): the upstream card used by both `competitive_coding_python` and `competitive_coding_cpp` lists `cc-by-4.0`, `odc-by`, and additional `mit` notice. The card states that the dataset is ready for commercial use. No dedicated citation block is provided on the card. - [nvidia/Nemotron-SFT-SWE-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-SWE-v2): the upstream card used by the `agentless` subset lists `cc-by-4.0` with additional `apache-2.0`, `mit`, `bsd-3-clause`, and `bsd-2-clause` notices. No dedicated citation block is provided on the card. - [nvidia/Nemotron-SFT-Instruction-Following-Chat-v2](https://huggingface.co/datasets/nvidia/Nemotron-SFT-Instruction-Following-Chat-v2): the upstream card used by the `reasoning_on` subset lists `odc-by`. No dedicated citation block is provided on the card. ## Intended Use This dataset is intended for supervised fine-tuning or continued instruction tuning of code-capable models, especially models targeting Python competitive programming and code reasoning. The auxiliary C++, SWE, and reasoning-chat slices are included to improve coverage rather than to define the core distribution. ## Limitations This is a mixed, processed, and partially synthetic dataset. It may inherit model-generated artifacts, reasoning mistakes, formatting noise, and licensing constraints from the upstream sources. It is better suited for training than for source-pure evaluation.