Datasets:
metadata
license: apache-2.0
task_categories:
- text-generation
Used Source Datasets:
- Heimdall-v1.1 Cybersecurity Dataset (Conversations)
- LiteCoder-Terminal-SFT
- Creative-Writing-High-Quality-1300x
Data Cleaning & Quality Gate Applied:
This dataset has been rigorously filtered and normalized using a strict automated pipeline:
- Deduplication: Exact and normalized duplicates across all sources were removed using disk-backed database hashing.
- Text Repair: Fixed broken text encoding (mojibake) and unified Unicode normalization forms (NFC) via
ftfy. - Heuristic Quality Filtering: Removed low-quality records including trivial/too short user prompts, abnormally short or cut-off assistant responses, and text loops (excessive character/word repetition).
- Privacy Gate: Screened and removed records containing high-confidence secrets (API keys, credentials).