LogicMind-Chat-Reasoning-SFT-300K

# Nemotron-Post-Training-Dataset-v2-chat Dataset Card ## Overview 📌 This dataset contains **296,168** chat-style instruction/response samples generated by `qwen-3-32b`. Each record provides a user prompt, an explicit reasoning trace, and a final answer, plus precomputed length fields. The data is packaged as JSONL (one JSON object per line). **Highlights** - **Scale:** 296,168 samples - **Category:** `chat` (100%) - **Generator:** `qwen-3-32b` (100%) - **Structure:** `problem` → `qwen3-reasoning` → `qwen3-solution` with length metadata - **Coverage:** Multi-domain instructions and multilingual content (based on random sampling) ## Dataset Structure 🧩 ### Fields | Field | Type | Description | | --- | --- | --- | | `id` | string | Unique record identifier | | `generator` | string | Model tag for data generation (`qwen-3-32b`) | | `category` | string | Data category (`chat`) | | `problem` | string | User prompt or instruction | | `qwen3-reasoning` | string | Chain-of-thought style reasoning trace | | `qwen3-solution` | string | Final answer to the prompt | | `len_qwen3_reasoning` | int | Length of reasoning text (provided) | | `len_qwen3_solution` | int | Length of final answer text (provided) | ### Example (schema) ```json { "id": "...", "generator": "qwen-3-32b", "category": "chat", "problem": "...", "qwen3-reasoning": "...", "qwen3-solution": "...", "len_qwen3_reasoning": 123, "len_qwen3_solution": 456 } ``` ## Statistical Summary 📊 **Total samples:** 296,168 **Length statistics (characters / length values)** | Field | Median | Mean | P95 | Max | | --- | --- | --- | --- | --- | | `problem` | 207 | 724.72 | 3,488 | 56,088 | | `qwen3-reasoning` | 701 | 914.63 | 2,499 | 8,333 | | `qwen3-solution` | 761 | 768.62 | 1,526 | 9,700 | **Answer-to-reasoning length ratio** (only for `len_qwen3_reasoning > 0`) - Median: 0.93 - P95: 2.49 ## Visualizations 🧪 - **Reasoning vs. solution lengths:**    ## Intended Uses 💡 - **Instruction tuning** for chat and multi-domain assistants - **Chain-of-thought distillation** or reasoning supervision - **Complex QA** and multi-step problem solving - **Benchmarking** reasoning length control and response compression ## Data Characteristics 🔎 - The dataset is **fully synthetic** (model-generated), not human-authored. - The reasoning traces are explicit and may include sensitive or unnecessary intermediate steps for deployment in certain applications. - The distribution of languages and domains is **non-uniform**; for downstream tasks, consider stratified sampling or domain filtering. ## Limitations ⚠️ - **Hallucination risk:** model-generated data may contain factual errors. - **Reasoning leakage:** the presence of explicit reasoning may be undesirable for certain production use cases. - **Domain bias:** real-world coverage is not guaranteed across all subject areas. ## Data Splits 🧩 - No official train/validation/test split is provided. Users should create their own splits as needed. ## License & Usage 👀 - Please follow the license terms associated with the source repository or data release. - If you use this dataset, cite the original source and acknowledge the generation model. ## Citation 🙏 ```bibtex @misc{nvidia2025nemotronv2, title = {Nemotron-Post-Training-Dataset-v2}, author = {NVIDIA}, year = {2025}, note = {Includes SFT and RL data across math, code, reasoning, and chat tasks}, } ``` ```bibtex @misc{Jackrong2026dataset, title = {LogicMind-Chat-Reasoning-SFT-300K}, author = {Jackrong}, year = {2026}, note = {A dataset constructed using seed questions inspired by NVIDIA's Nemotron-Post-Training-Dataset-v2-chat} } ```许可证：Apache-2.0 语言： - 英语 ---

搜集汇总

背景与挑战

背景概述

该数据集包含约30万条由qwen-3-32b生成的聊天式指令/响应样本，每个样本提供用户提示、显式推理过程和最终答案，以JSONL格式存储。它适用于指令微调、推理监督和多步问题解决等任务，但需注意数据为合成生成，可能存在事实错误或领域偏差风险。