SKU SFT-2108 · Sold by External
Dolphin
Product specifications
| SKU | SFT-2108 |
|---|---|
| Data type | SFT dataset |
| Volume | 3731947 examples |
| Size on disk | 18.9 GB (source JSONL) |
| Format | JSONL (Parquet auto-conversion available) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | Apache-2.0 |
Dolphin is Eric Hartford's open reproduction of Microsoft's Orca, built on the FLAN v2 task mixture. It ships two configs: flan1m-alpaca-uncensored (891,857 rows, GPT-4 completions) and flan5m-alpaca-uncensored (2,840,090 rows, GPT-3.5 completions), ~3.73M examples total. Instructions are augmented with model completions and filtered to remove alignment, refusal, and bias content ('uncensored'). Rows use the Alpaca instruction/input/output schema. The dataset now lives at QuixiAI/dolphin (the cognitivecomputations/dolphin path redirects there).