---
license: odc-by

configs:
- config_name: default
  data_files:
    - split: "0.7"
      path: "prompts_e0p700.parquet"
    - split: "0.75"
      path: "prompts_e0p750.parquet"
    - split: "0.8"
      path: "prompts_e0p800.parquet"
    - split: "0.9"
      path: "prompts_e0p900.parquet"
---

These are a collection of deduplicated English prompts shorter than ~2000 tokens taken from WildChat-4.8M.
- Initially, many non-English prompts were removed by validating that at least 80% of each prompt was composed of characters from the English alphabet.
- Then, a HashSet was used to directly deduplicate identical prompts (ignoring whitespace and punctuation).
- Then, MinHash was used to further conservatively deduplicate.
- Finally, Qwen3-8B-Embedding was used to generate embeddings on all remaining prompts. They were clustered according to a certain similarity threshold, and only the prompt furthest from the centroid was kept while the rest were removed.
  -  The resulting datasets are provided at four similarity thresholds used for clustering: 0.7, 0.75, 0.8, and 0.9.

```
Declustering:
eps=0.700 -> kept 232,725 / 655,426 (35.51%)
eps=0.750 -> kept 297,955 / 655,426 (45.46%)
eps=0.800 -> kept 358,697 / 655,426 (54.73%)
eps=0.900 -> kept 476,493 / 655,426 (72.70%)
```

A sample of 1000 clusters that were removed per each similarity level is provided in the repo.

Prompts longer than 2000 tokens will be added at a later date.

**Note: The 0.8 split here has different results than the preceding dataset at https://huggingface.co/datasets/MasonMac/WildChat-4M-English-Semantic-Deduplicated because it switched from connected components to "Greedy Algorithm for Maximum Independent Set." This idea was stolen from SemDeDup to mitigate false positives. Other changes were made as well.**