dkoplow/ZKL-hh

---
pretty_name: ZKL HH (parquet)
tags:
- rlhf
- hh
---

# ZKL HH (parquet)

This dataset contains the `hh` split exported by `gen_hf_ds.py`.
Rows are deduplicated and split into `train` and `val`, with `val` containing
2000 examples. You must comply with the licenses and terms of the upstream
dataset.

## Layout

- `train/data/train.parquet`
- `val/data/val.parquet`

## Columns

- `dataset`: always `"hh"`
- `prompt`: chat list of `{"role", "content"}` dicts
- `label`: Always `null` for HH.
- `id`: stable string id

## Load with `datasets`

```python
from datasets import load_dataset

repo = "dkoplow/ZKL-hh"
train = load_dataset(
"parquet",
data_files=f"hf://datasets/{repo}/train/data/train.parquet",
split="train",
)
val = load_dataset(
"parquet",
data_files=f"hf://datasets/{repo}/val/data/val.parquet",
split="train",
)
```

---
数据集名称：ZKL HH（Parquet格式）
标签：
- 强化学习人类反馈（Reinforcement Learning from Human Feedback）
- hh
---

# ZKL HH（Parquet格式）

本数据集包含由`gen_hf_ds.py`导出的`hh`划分数据。所有样本行已完成去重，并划分为训练集（train）与验证集（val），其中验证集包含2000条样本。使用者需遵守上游数据集的许可协议与使用条款。

## 数据布局

- `train/data/train.parquet`
- `val/data/val.parquet`

## 字段说明

- `dataset`：固定取值为`"hh"`
- `prompt`：由包含`"role"`和`"content"`键的字典组成的对话列表
- `label`：针对HH数据集，该字段始终为`null`（空值）
- `id`：稳定字符串标识符

## 使用`datasets`库加载

python
from datasets import load_dataset

repo = "dkoplow/ZKL-hh"
train = load_dataset(
"parquet",
data_files=f"hf://datasets/{repo}/train/data/train.parquet",
split="train",
)
val = load_dataset(
"parquet",
data_files=f"hf://datasets/{repo}/val/data/val.parquet",
split="train",
)