xlr8harder's picture
Update README with v2 cleaning stages
96357ec verified
|
Raw
History Blame Contribute Delete
4.25 kB
metadata
license: odc-by
language:
  - en
tags:
  - wildchat
  - filtered
  - prompts
  - training-data
  - openness-rated
pretty_name: WildChat Filtered Rated Prompts
size_categories:
  - 100K<n<1M

WildChat Filtered Rated Prompts

A cleaned subset of the WildChat dataset, filtered for quality and rated for "prompt openness" (how much a prompt invites genuine engagement vs mechanical task completion).

Dataset Description

305,246 unique user prompts with openness ratings (1-5 scale).

Score Description Count
5 Highly open — invites reflection, opinion, dialogue 40,792
4 Mostly open — flexible approach, room for voice 52,900
3 Mixed — could go either way 15,081
2 Mostly closed — limited flexibility 118,932
1 Fully closed — single correct answer 77,541

Format

JSONL with one prompt per line:

{"id": "sha256hash", "prompt": "user prompt text", "score": 4, "analysis": "Brief explanation of rating"}

Pipeline

Stage 1: Initial Filtering (329k → 305k)

From xlr8harder/wildchat-filtered-prompts:

Stage Output Method
Raw WildChat 3.1M Original dataset
Extract + dedup 2.4M First user message, SHA256 dedup
Length + language 962k 20-5000 chars, English (langdetect)
Fuzzy dedup 456k MinHash LSH, Jaccard 0.7
N-gram spam 351k Trigram frequency filtering
Cluster spam 329k TF-IDF + KMeans tight clusters

Stage 2: Openness Rating (329k)

Each prompt rated 1-5 for "openness" using LLM classification with the following rubric:

  • 5 (Highly Open): Invites reflection, opinion, creative exploration, genuine dialogue
  • 4 (Mostly Open): Room for interpretation, flexible approach, voice matters
  • 3 (Mixed): Could be answered mechanically or with engagement
  • 2 (Mostly Closed): Specific task, limited flexibility, mostly one right approach
  • 1 (Fully Closed): Single correct answer, pure lookup/computation

Stage 3: Additional Cleaning (329k → 305k)

Filter Removed Method
Non-English 15,649 (4.8%) 10+ non-Latin Unicode characters
Template spam 8,368 (2.5%) Prefix-based dedup, cap at 2 per template

Non-English removal: Prompts with significant non-Latin script (Cyrillic, Chinese, Arabic, Korean, Japanese, Hebrew, Thai, Devanagari, Tamil, Telugu, Bengali). Note: langdetect in Stage 1 caught most non-English, but some mixed-language prompts remained.

Template spam detection: Identified 339 template patterns (prompts sharing the same 50-char prefix, appearing 10+ times). Each template capped at 2 examples for diversity. 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 scam-baiting response to the following email..." (185 → 2)
  • "Create a hypothetical fact sheet/history/commentary..." (~900 → 6)
  • Various Russian/French/Vietnamese templates caught by non-English filter

Intended Use

  • Training data for language models (diverse prompt types)
  • Research on user prompts and interaction patterns
  • Studying prompt diversity and openness characteristics

Limitations

  • English only (non-English filtered out)
  • First turn only (multi-turn context not preserved)
  • Openness ratings are LLM-generated, not human-verified
  • Template detection may have false positives/negatives

Source

License

ODC-BY (inherited from WildChat)

Citation

@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}
}