promptcot
PromptCoT-2.0-SFT-4.8M
PromptCoT-2.0-SFT-4.8M
This repository contains the largest dataset released with PromptCoT 2.0 (Scaling Prompt Synthesis for LLM Reasoning).It includes 4.8 million fully synthetic prompts with reasoning trajectories, serving as the cornerstone for supervised fine-tuning (SFT) experiments.
The dataset demonstrates that purely synthetic data—when generated with PromptCoT 2.0—can train competitive reasoning models that outperform human-curated baselines such as OpenMathReasoning and… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-2.0-SFT-4.8M.PromptCoT-2.0-SelfPlay-4B-48K
PromptCoT-2.0-SelfPlay Datasets
This repository hosts the self-play datasets used in PromptCoT 2.0 (Scaling Prompt Synthesis for LLM Reasoning).These datasets were created by applying the PromptCoT 2.0 synthesis framework to generate challenging math and programming problems, and then training models through self-play with Direct Preference Optimization (DPO).
PromptCoT-2.0-SelfPlay-4B-48K: 48,113 prompts for Qwen3-4B-Thinking-2507 self-play.
PromptCoT-2.0-SelfPlay-30B-11K: 11… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-2.0-SelfPlay-4B-48K.PromptCoT-DS-Dataset
Dataset Format
Each row in the dataset contains:
prompt: The input to the reasoning model, including a problem statement with the special prompting template.
completion: The expected output for supervised fine-tuning, containing a thought process wrapped in <think>...</think>, followed by the final solution.
Example
{
"prompt": "<|begin▁of▁sentence|>Please reason step by step, and put your final answer within \\boxed{{}}.<|User|>\nLet $P$ be a point on a… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-DS-Dataset.PromptCoT-2.0-Concepts
🧩 PromptCoT 2.0 – Concepts Dataset
PromptCoT 2.0 Concepts provides the foundational conceptual inputs for large-scale prompt synthesis in mathematics and programming.These concept files are used to generate high-quality synthetic problems through the PromptCoT 2.0 Prompt Generation Model.
📘 Overview
Each file (e.g., math.jsonl, code.jsonl) contains a list of concept prompts that serve as the input for the problem generation stage.By feeding these prompts into the… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-2.0-Concepts.PromptCoT-2.0-SelfPlay-30B-11K
PromptCoT-2.0-SelfPlay Datasets
This repository hosts the self-play datasets used in PromptCoT 2.0 (Scaling Prompt Synthesis for LLM Reasoning).These datasets were created by applying the PromptCoT 2.0 synthesis framework to generate challenging math and programming problems, and then training models through self-play with Direct Preference Optimization (DPO).
PromptCoT-2.0-SelfPlay-4B-48K: 48,113 prompts for Qwen3-4B-Thinking-2507 self-play.
PromptCoT-2.0-SelfPlay-30B-11K: 11… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-2.0-SelfPlay-30B-11K.PromptCoT-Problem-Generation-Dataset
Dataset Format
Each row in the dataset contains:
prompt: The input to the problem generation model, including foundational concepts and the required difficulty level (e.g., AIME, HMMT-Feb).
completion: The expected output for supervised fine-tuning, consisting of a rationale (detailing the problem design process) and the final mathematical problem.
Example
{
"prompt": "Given foundational concepts and difficulty level, identify connections and develop a… See the full description on the dataset page: https://huggingface.co/datasets/xl-zhao/PromptCoT-Problem-Generation-Dataset.