SciRIFF-train-mix

# SciRIFF training mix

This dataset includes the training mix used to train the SciTulu models described in our paper [SciRIFF: A Resource to Enhance Language Model Instruction-Following over Scientific Literature](https://arxiv.org/abs/2406.07835).

It contains 35K instances from the [SciRIFF](https://huggingface.co/datasets/allenai/SciRIFF) dataset (1,000 instances per train task), together with a matching number of instances randomly sampled from the [Tulu V2 mix](https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture). See the dataset cards of the component datasets for more information.

You can load the dataset like:

```python
import datasets
ds = datasets.load_dataset("allenai/SciRIFF-train-mix")
```

The instances are formatted the same way as in the Tulu V2 mix, with the following fields:

- `dataset`: Identifier for the dataset this instance belongs to.
- `id`: Unique ID for the instance. For SciRIFF instances, this is the same as the `_instance_id` field in the [SciRIFF dataset](https://huggingface.co/datasets/allenai/SciRIFF).
- `messages`: A list of messages, for instance:

```python
[
{"role": "user", "content": [user message]},
{"role": "assistant", "content": [model response]}
]
```

For more information on how to train models using this dataset, see our GitHub repository: https://github.com/allenai/SciRIFF.

# SciRIFF训练混合数据集

本数据集为论文《SciRIFF：提升大语言模型在科学文献上的指令遵循能力的资源》（https://arxiv.org/abs/2406.07835）中所介绍的SciTulu模型的训练所用的训练混合数据集。

本数据集包含3.5万条来自SciRIFF数据集的样本（每个训练任务包含1000条样本），同时搭配从Tulu V2混合数据集（https://huggingface.co/datasets/allenai/tulu-v2-sft-mixture）中随机采样的等量样本。更多详情请参阅各组分数据集的数据集卡片文档。

可通过以下代码加载该数据集：

python
import datasets
ds = datasets.load_dataset("allenai/SciRIFF-train-mix")

该数据集的样本格式与Tulu V2混合数据集一致，包含以下字段：

- `dataset`：该样本所属数据集的标识符。
- `id`：该样本的唯一标识符。对于SciRIFF样本而言，其值与SciRIFF数据集（https://huggingface.co/datasets/allenai/SciRIFF）中的`_instance_id`字段一致。
- `messages`：消息列表，示例如下：

python
[
{"role": "user", "content": [user message]},
{"role": "assistant", "content": [model response]}
]

如需了解使用该数据集训练模型的更多详情，请参阅我们的GitHub仓库：https://github.com/allenai/SciRIFF。