tags:

• anthropic
• claude
• hh-rlhf
• prompts
• wildchat
• instruction-data
• preference-data
• raw-data
• merged-dataset

Anthropic.Chat

Important: this repository is not official Anthropic software or an official Anthropic dataset. It is a derived merged dataset created from public Hugging Face sources.

Subset details

• Max rows per source dataset: None
• Max total rows: None
• Total rows written: 179,315
• Parquet shards: 360
• Approximate local size: 281.34 MB
• Shard rows: 500
• Include extracted fields: True

Row counts

• princeton-nlp/rl_tulu3_wildchat-if_prompts: 7,794 rows
• fka/prompts.chat: 2,169 rows
• Anthropic/hh-rlhf: 169,352 rows

Splits/configs used

• princeton-nlp/rl_tulu3_wildchat-if_prompts splits/configs: train|config=default: 7,794
• fka/prompts.chat splits/configs: train|config=default: 2,169
• Anthropic/hh-rlhf splits/configs: train|config=default: 160,800, test|config=default: 8,552

Merge errors

• none

Schema

This merged dataset preserves the full original row in:

sample["raw_row_json"]

Additional metadata columns:

If extracted fields are enabled, these columns are also included:

Why raw_row_json?

The source datasets have different schemas.

To avoid schema errors and preserve all original information, this merged dataset stores the full original row as JSON:

import json

original_row = json.loads(sample["raw_row_json"])

Usage

from datasets import load_dataset
import json

ds = load_dataset(
    "antontuzovAI/Anthropic.Chat",
    streaming=True,
    split="train"
)

for sample in ds:
    print(sample["source_dataset"])
    print(sample["source_split"])
    print(sample["source_config"])

    original_row = json.loads(sample["raw_row_json"])
    print("Original keys:", list(original_row.keys()))
    break

Preference-style usage

For rows from Anthropic/hh-rlhf, you may often have:

sample["chosen"]
sample["rejected"]

Example:

from datasets import load_dataset

ds = load_dataset(
    "antontuzovAI/Anthropic.Chat",
    streaming=True,
    split="train"
)

for sample in ds:
    if sample["source_dataset"] == "Anthropic/hh-rlhf":
        print("CHOSEN:", sample["chosen"][:500])
        print("REJECTED:", sample["rejected"][:500])
        break

Prompt-style usage

For rows from fka/prompts.chat or princeton-nlp/rl_tulu3_wildchat-if_prompts, you may often have:

sample["prompt"]
sample["title"]
sample["category"]
sample["url"]

Example:

from datasets import load_dataset

ds = load_dataset(
    "antontuzovAI/Anthropic.Chat",
    streaming=True,
    split="train"
)

for sample in ds:
    if sample["source_dataset"] == "fka/prompts.chat":
        print("TITLE:", sample["title"])
        print("CATEGORY:", sample["category"])
        print("PROMPT:", sample["prompt"][:500])
        print("URL:", sample["url"])
        break

Source datasets

This dataset is derived from:

• princeton-nlp/rl_tulu3_wildchat-if_prompts
• fka/prompts.chat
• Anthropic/hh-rlhf

License and usage notice

You must respect the license and terms of each source dataset.

This merged dataset does not grant rights beyond the original source licenses.

If you use this data for training or evaluation, review the source dataset cards carefully.