vibe_coding_900k

# Full Vibe Coding 898109 This is the portable, globally deduplicated training-input pool assembled from the 719,061-row expanded main vibe-coding dataset and the 200,000-row agentic supplement. It contains 898,109 unique rows: - expanded main pool: 719,061 - retained agentic supplement: 179,048 - overlapping agentic source records removed: 20,952 When the two components contain the same source record, the expanded-main version is retained because it carries the richer repository and validation context. The separate Chinese supplement is not included in this package. Rows contain prompts, input context, repository provenance and per-row license metadata. Source assistant answers and gold patches are excluded. See `manifest.json` for the schema and composition, and `validation.json` for duplicate checks and shard hashes. ## Load with Hugging Face Datasets ```python from datasets import load_dataset train = load_dataset("YOUR_NAMESPACE/YOUR_DATASET", split="train") ``` A downloaded snapshot can also be loaded directly: ```python train = load_dataset( "parquet", data_files="data/train-*.parquet", split="train" ) ``` ## ModelScope Upload the whole directory as one dataset repository. It uses plain Parquet shards under `data/` and does not require a custom loading script. ## License This is a multi-source collection, so there is no single source license. Consult each row's `license`, `source`, and `source_id` fields before use.