license: mit
task_categories:

• text-generation
• reinforcement-learning
  language:
• en
  tags:
• grpo
• rlhf
• cleaned
• deduplicated
• reward-model
• reasoning
• code
• safety
  size_categories:
• 1K<n<10K
  source_datasets:
• Anthropic/hh-rlhf

HH-RLHF-Cleaned-GRPO

📋 Dataset Description

Safety and helpfulness prompts extracted from Anthropic's HH-RLHF dataset. This cleaned version removes 100 duplicate prompts and 87 whitespace issues that would waste training compute on redundant data.

Original source: Anthropic/hh-rlhf by Anthropic

📊 Cleaning Statistics

Bug Breakdown by Type

🔍 Bug Records (Line-by-Line)

Every bug found during cleaning is recorded below with line number, type, and action taken.
This provides a verifiable audit trail of all changes made.

🧹 Cleaning Methodology

This dataset was cleaned using a 14-point deep inspection system:

📝 Data Format

Each line is a JSONL object:

{"prompt": "Write a function to...", "type": "code_critique"}

💻 Usage

from datasets import load_dataset

# Load the cleaned dataset
ds = load_dataset("Eyght/HH-RLHF-Cleaned-GRPO", split="train")

# Use for GRPO training
for example in ds:
    prompt = example["prompt"]
    # Generate completions and score with reward functions

✅ Verification

• Remaining bugs after cleaning: 0
• All prompts verified: ≥10 chars, valid JSON, no HTML, no control chars
• No duplicates: All prompts unique (case-insensitive)
• Encoding: UTF-8 throughout
• License: mit (same as original)

🏷️ Citation

If you use this cleaned dataset:

@dataset{eyght_HH-RLHF-Cleaned-GRPO_2026,
  title     = {HH-RLHF-Cleaned-GRPO},
  author    = {Eyght},
  source    = {Anthropic/hh-rlhf},
  year      = {2026},
  month     = {September},
  note      = {Deep-cleaned: 187 bugs fixed from 1947 original examples},
  url       = {https://huggingface.co/datasets/Eyght/HH-RLHF-Cleaned-GRPO}
}

📜 License

mit — same as the original dataset. Free for commercial and non-commercial use.

Cleaned by Eyght | September 10, 2026

This dataset card includes a complete audit trail of all bugs found and fixed. Every line-level change is recorded in the Bug Records table above for full transparency and reproducibility.