LingReason
收藏资源简介:
LingReason Chintang数据集是LingReason项目的一部分,专门用于低资源机器翻译研究,特别是探索语言学推理轨迹对翻译性能的影响。该数据集包含Chintang语(一种低资源语言)的翻译数据,由LingReason GitHub仓库发布的代码生成,并伴随研究论文《Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?》。数据集包含六个分割:训练集(1,831个样本,包含完整语言学推理轨迹)、验证集(114个样本,包含完整语言学推理轨迹)、测试集(344个样本,包含完整语言学推理轨迹)、无推理轨迹训练集(1,831个样本)、无推理轨迹验证集(114个样本)以及用于上下文学习实验的测试集(344个样本,提示中包含带占位符的语言学推理引导)。数据集设计用于比较有/无语言学推理辅助的翻译模型性能,适用于监督微调、强化微调和上下文学习等机器翻译实验场景。
The LingReason Chintang Dataset is part of the LingReason project, specifically dedicated to low-resource machine translation research, particularly exploring the impact of linguistic reasoning traces on translation performance. This dataset contains translation data for the Chintang language (a low-resource language), generated by the code released in the LingReason GitHub repository, and is accompanied by the research paper titled "Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?". The dataset includes six splits: the training set (1,831 samples with complete linguistic reasoning traces), the validation set (114 samples with complete linguistic reasoning traces), the test set (344 samples with complete linguistic reasoning traces), the training set without reasoning traces (1,831 samples), the validation set without reasoning traces (114 samples), and the test set for in-context learning experiments (344 samples with placeholder-integrated linguistic reasoning guidance in the prompts). This dataset is designed to compare the translation performance of models with and without linguistic reasoning assistance, and is applicable to machine translation experimental scenarios such as supervised fine-tuning, reinforcement fine-tuning, and in-context learning.
数据集概述:LingReason Chintang Data
- 来源:由 LingReason 项目生成,代码发布在 LingReason GitHub 仓库,并伴随论文《Reasoning over Grammar: Can Synthetic Linguistic Reasoning Traces Enhance Low-Resource Machine Translation?》(arXiv:2606.03782)。
- 许可协议:cc-by-nc-sa-4.0
- 语言:ctn(Chintang 语)
- 任务类型:翻译(translation)
- 标签:低资源机器翻译(low-resource-machine-translation)、语言推理(linguistic-reasoning)、通用依存关系(universal-dependencies)、Chintang(chintang)
数据集划分与文件
数据集包含 6 个划分,每个划分对应一个 JSON 文件:
| 划分 | 文件 | 样本数 | 描述 |
|---|---|---|---|
test_icl |
ctn_test_icl.json |
344 | 测试集,提示中包含带占位符的语言推理指南,用于上下文学习实验 |
train |
ctn_train.json |
1,831 | SFT/RFT 训练集,在 <think> 块中包含完整的语言推理轨迹 |
validation |
ctn_eval.json |
114 | 评估/验证集,在 <think> 块中包含完整的语言推理轨迹 |
train_no_thinking |
ctn_no_thinking_train.json |
1,831 | SFT 训练集,不包含语言推理轨迹 |
validation_no_thinking |
ctn_no_thinking_eval.json |
114 | 评估/验证集,不包含语言推理轨迹 |
test |
ctn_test.json |
344 | 测试集,在 <think> 块中包含完整的语言推理轨迹,用于 ICL 基线以及 SFT 和 RFT 实验 |
使用示例
可通过 Hugging Face datasets 库加载:
python from datasets import load_dataset
dataset = load_dataset("OLAResearchX/LingReason") print(dataset) print(dataset["train"][0])




