X-Coder-SFT-376k

# X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests ## Dataset Overview X-Coder-SFT-376k is a large-scale, fully synthetic dataset for advancing competitive programming. The dataset comprises **4 subsets** with a total of **887,321 synthetic records** across **423,883 unique queries**. It is designed for supervised fine-tuning and suitbale for cold start to train code reasoning foundations. X-Coder-SFT-376k is curated by sota reasoning models. The query is synthesized by [OpenAI GPT-o3-mini](https://openai.com/zh-Hant/index/openai-o3-mini/), the solutions are generated by [DeepSeek-R1-0528](deepseek-ai/DeepSeek-R1-0528) and [Qwen3-235B-A22B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Thinking-2507). [Technical Report](https://arxiv.org/abs/2601.06953) - Discover the methodology details behind the Dataset. [Github Repo](https://github.com/JieWu02/X-Coder) - Access the complete pipeline to generate your own artifacts, and perform SFT & RLVR. ## Data Scaling  Scaling laws on the generated SFT dataset. Left: Performance comparison of on LiveCodebench v5 to examine scaling trend. Right: Performance comparison across scaling unique tasks and scaling solutions per task. ## Dataset Statistics <h4 style="margin-bottom: 0px; padding-bottom: 0px;">hybrid</h4> | Responses | Queries | Percentage | |-----------|---------|------------| | 1 | 142,404 | 80.48% | | 2-7 | 17,329 | 9.79% | | 8 | 13,968 | 7.89% | | 9-16 | 3,234 | 1.83% | #### unique-prompt - 100% queries with 1 response #### verified - 100% queries with exactly 8 responses - All verified high-quality solutions <h4 style="margin-bottom: 0px; padding-bottom: 0px;">multiple-solution</h4> | Responses | Queries | Percentage | |-----------|---------|------------| | 1 | 89 | 0.27% | | 2 | 328 | 0.98% | | 3 | 1,403 | 4.18% | | 4 | 2,339 | 6.97% | | 5 | 4,292 | 12.78% | | 6 | 5,653 | 16.84% | | 7 | 6,738 | 20.07% | | 8 | 12,729 | 37.92% | <h4 style="margin-bottom: 0px; padding-bottom: 0px;">Record Distribution</h4> | Split | Records | Unique Queries | Responses/Query | Thinking Tags | Size | |-------|---------|----------------|-----------------|----------------|------| | **unique-prompt** | 202,128 | 202,125 | 1.00 | ❌ No | 13.1 GB | | **multiple-solution** | 218,686 | 33,571 | 6.51 | ✅ Yes | 13.6 GB | | **hybrid** | 376,491 | 176,935 | 2.13 | ✅ Yes | 24.3 GB | | **verified** | 90,016 | 11,252 | 8.00 | ✅ Yes | 5.2 GB | ## Dataset Splits ### 1. `hybrid` (Hybrid-SFT-376k) - **Purpose**: Large-scale SFT training - **Features**: - Mixed dataset with three tiers: - 80.48% single-response queries - 7.89% queries with 8 verified responses - 11.63% queries with 2-16 responses - 100% include reasoning process by R1-0528 and Qwen3-235b-a22-2507-thinking - Contains the entire `verified` split - Partially overlaps with `multiple-solution` (~55%) ### 2. `unique-prompt` (SFT-unique-prompt-202k) - **Features**: - Better query diversity. Each query has exactly 1 response - Completely independent from other splits ### 3. `multiple-solution` (SFT-multiple-solution-218k) - **Purpose**: Training models to generate diverse solutions - **Features**: - Better solution diversity. Most queries with multiple sampled responses (avg 6.51) - Response count distribution: 1-8 responses per query - **Use Cases**: Solution diversity ### 4. `verified` (SFT-verified-90k) - **Purpose**: High-quality verified solutions - **Features**: - Exactly 8 verified responses per query - Subset of `hybrid` split (100% contained) ## Data Format Each record is a JSON object with the following structure: ```json { "query": "Problem description or coding challenge...", "response": "Solution with optional <think>reasoning process</think> and answer..." } ``` ## Dataset Relationships ``` unique-prompt (202k) is completely independent with Other splits multiple-solution (218k) ↓ (~55% overlap) hybrid (376k) ↓ (100% contained) verified (90k) ``` ### Overlap: - `unique-prompt` ∩ others = ∅ (no overlap) - `multiple-solution` ∩ `hybrid` ≈ 55% - `verified` ⊂ `hybrid` (100% contained) - `multiple-solution` ∩ `verified` = ∅ If mixing multiple splits: - ✅ `unique-prompt` + any other: No deduplication needed - ⚠️ `multiple-solution` + `hybrid`: Requires deduplication (~55% overlap) - ⚠️ `verified` + `hybrid`: Use `hybrid` only (verified is fully contained) ## Usage Examples ```python # Use the comprehensive hybrid split SFT-376k = load_dataset("IIGroup/X-Coder-SFT-376k", split="hybrid") ``` ### 📝 Citation ``` @article{wu2026x, title={X-Coder: Advancing Competitive Programming with Fully Synthetic Tasks, Solutions, and Tests}, author={Wu, Jie and Li, Haoling and Zhang, Xin and Guo, Jiani and Luo, Jane and Liu, Steven and Huang, Yangyu and Chu, Ruihang and Li, Scarlett and Yang, Yujiu}, journal={arXiv preprint arXiv:2601.06953}, year={2026} } ``` ## License This dataset is licensed under the MIT license. --- **Version**: 2026-01-05