🐋 Orca-AgentInstruct-1M-v1-cleaned

This is a cleaned version of the microsoft/orca-agentinstruct-1M-v1 dataset released by Microsoft.

orca-agentinstruct-1M-v1 is a fully synthetic dataset using only raw text publicly available on the web as seed data. It is a subset of the full AgentInstruct dataset (~25M samples) that created Orca-3-Mistral. Compared to Mistral 7B Instruct, the authors claim 40% improvement on AGIEval, 19% improvement on MMLU, 54% improvement on GSM8K, 38% improvement on BBH and 45% improvement on AlpacaEval.

Here's what I changed:

1. Splits are unified into one, with a new "split" column
2. Strings were converted into lists of dicts to ensure compatibility with most frameworks
3. Empty system prompts were removed so you don't get weird errors

Data categories in the dataset:

• creative_content
• text_modification
• struct2text_flow
• rc
• rag
• text_extraction
• mcq
• follow_up
• analytical_reasoning
• fermi
• fs_cot_flow
• code_
• brain_teaser
• text_classification
• open_domain_q