aurora-m2 aka safellm

Synthetic Data Generation Papers

• Creating more data from smaller models is more efficient - Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling.
• Evol-Instruct - Automatic Instruction Evolving for Large Language Models
• Orca-3 - AgentInstruct: Toward Generative Teaching with Agentic Flows
• 1 Billion Personas - Scaling Synthetic Data Creation with 1,000,000,000 Personas Code
• SAT - SATLM: Satisfiability-Aided Language Models Using Declarative Prompting
• Arena Learning: Build Data Flywheel for LLMs Post-training via Simulated Chatbot Arena
• Magpie - Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing

Tools/ Library

Benchmark Report of Different LLM Inference Backends link

• LMDeploy [Fastest]
• vLLM

Installation

• On leonardo, you will need to do this: module load cuda
• pip install -r requirements.txt
• python -m wn download oewn:2023
• python -m spacy download en_core_web_sm
• python -m spacy download xx_ent_wiki_sm
• NOTE: scispacy is not working on Leonardo for some reason bc of nmslib not installing. TODO to fix.
• pip install -q https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.5.4/en_ner_bc5cdr_md-0.5.4.tar.gz