xlr8harder/wildchat-filtered-prompts

WildChat Filtered Prompts

A cleaned and deduplicated subset of the WildChat dataset, filtered for quality and diversity.

Dataset Description

305,246 unique English user prompts extracted from WildChat, processed through a multi-stage pipeline to remove duplicates, spam, non-English content, and template spam.

Source

• Original Dataset: allenai/WildChat (GPT-4/ChatGPT user conversations)
• Original Size: 3,185,010 conversation turns
• Final Size: 305,246 prompts (90% reduction)

Pipeline

Stage	Output	Method
0. Raw WildChat	3,185,010	Parquet files
1. Extract	2,379,985	First user message only, SHA256 exact dedup
2. Length + Language	961,882	20-5000 chars, English only (langdetect)
3. Fuzzy Dedup	455,615	MinHash LSH, Jaccard threshold 0.7
4. N-gram Spam Filter	350,970	Trigram doc frequency >0.5% indicates template spam
5. Cluster Spam Filter	328,875	TF-IDF + KMeans, remove tight near-identical clusters
6. Non-English Filter	313,226	Remove prompts with 10+ non-Latin Unicode chars
7. Template Dedup	305,246	Prefix-based template detection, cap at 2 per template

Stage Details

Stage 1 - Extraction: Extract only the first user message from each conversation. Many WildChat conversations are multi-turn, but the first message captures the user's original intent. Exact deduplication via SHA256 hash.

Stage 2 - Basic Filtering: Remove very short (<20 chars) and very long (>5000 chars) prompts. Filter to English using langdetect library.

Stage 3 - Fuzzy Deduplication: MinHash LSH to find near-duplicate prompts. Jaccard similarity threshold of 0.7 removes prompts that are substantially similar even if not exact matches.

Stage 4 - N-gram Spam Detection: Build trigram document frequency index. Trigrams appearing in >0.5% of all documents indicate template/spam patterns. Prompts with 2+ high-frequency trigram hits are removed.

Stage 5 - Cluster Spam Detection: TF-IDF vectorization + KMeans clustering. Identify "tight" clusters where prompts are very similar to centroid and to each other. These indicate coordinated spam or bot activity.

Stage 6 - Non-English Filter (Dec 2024): Remove prompts containing 10+ characters from non-Latin Unicode ranges (Cyrillic, Chinese, Arabic, Korean, Japanese, Hebrew, Thai, Devanagari, Tamil, Telugu, Bengali). This catches mixed-language prompts that langdetect missed.

Stage 7 - Template Deduplication (Dec 2024): Identify template patterns by looking for prompts with identical 50-character prefixes appearing 10+ times. Cap each template at 2 examples to maintain some diversity while removing spam. Top templates removed:

• "Write dialogue from a scene from the animated teen series Jane..." (268 → 2)
• "generate a description with keywords and SEOs for..." (249 → 2)
• "Create a hypothetical fact sheet/history/commentary..." (~900 → 6)
• Various single-user creative projects and SEO spam

Format

JSONL with one prompt per line:

{"id": "sha256hash", "prompt": "user prompt text"}

Derived Datasets

• xlr8harder/wildchat-filtered-rated-prompts — This dataset with openness ratings (1-5 scale)

Intended Use

• Training data for language models
• Research on user prompts and interactions
• Studying prompt diversity and patterns

Limitations

• English only (non-English prompts filtered out)
• First turn only (multi-turn context not preserved)
• Some false positives in spam detection (legitimate prompts may be removed)
• Some false negatives (spam patterns not caught by heuristics)

License

This dataset inherits the ODC-BY license from the original WildChat dataset.

Citation

If you use this dataset, please cite the original WildChat paper:

@article{zhao2024wildchat,
  title={WildChat: 1M ChatGPT Interaction Logs in the Wild},
  author={Zhao, Wenting and Ren, Xiang and Hessel, Jack and Cardie, Claire and Choi, Yejin and Deng, Yuntian},
  journal={arXiv preprint arXiv:2405.01470},
  year={2024}
}