Jayfeather1024/Reward-Embeddings-30k

---
license: unknown
---

# RLHF Reward Model Embedding Features for PKU-Alignment/PKU-SafeRLHF Dataset

The RLHF reward model embedding features and corresponding original text are stored in `embeddings_train_30k.jsonl` and `embeddings_test.jsonl`.

The dataset is stored in pairwise ways: each data pair has 1) safer_example: input text of the safer example, 2) not_safer_example: input text of the more harmful example, 3) safer_embedding: embedding feature of the safer example, 4) not_safer_embedding: embedding feature of the more harmful example.

The hidden embedding dimension is 4096. The reward model uses a linear layer to transfer the embedding features into a 1-dimensional score value.

Note: The dataset is extremely large because of the large size of the original training dataset and the high dimension of embedding space.

# Original Dataset

If you need more detailed information about the original dataset, please refer to `train.jsonl.xz` and `test.jsonl.xz`. Since we use `shuffle=False` when generating the embeddings, orders are remained in our dataset.

# Note

This dataset is a processed version of PKU-Alignment/PKU-SafeRLHF: <https://huggingface.co/datasets/PKU-Alignment/PKU-SafeRLHF>.

原始信息汇总

RLHF Reward Model Embedding Features for PKU-Alignment/PKU-SafeRLHF Dataset

数据集描述

• 文件存储：包含训练集 embeddings_train_30k.jsonl 和测试集 embeddings_test.jsonl。
• 数据结构：每个数据对包含以下内容：
  1. safer_example：更安全的输入文本。
  2. not_safer_example：更有害的输入文本。
  3. safer_embedding：更安全的示例的嵌入特征。
  4. not_safer_embedding：更有害的示例的嵌入特征。
• 嵌入维度：隐藏嵌入维度为 4096。
• 评分模型：奖励模型使用线性层将嵌入特征转换为 1 维的分数值。
• 数据集大小：由于原始训练数据集的规模较大和嵌入空间的高维度，数据集非常庞大。

原始数据集

• 详细信息：原始数据集的详细信息可参考 train.jsonl.xz 和 test.jsonl.xz。
• 数据顺序：生成嵌入时使用 shuffle=False，因此数据集中的顺序得以保留。

备注

• 数据集来源：该数据集是 PKU-Alignment/PKU-SafeRLHF 的加工版本。