EVE: An Open Source Earth Science LLM for Researchers, Policymakers, and the Public
summary
EVE fine-tunes Mistral 3.2 24B on a curated EO corpus (about 5.3B open tokens plus 1.1B closed tokens), creates ~1B synthetic QA tokens grounded in retrieved chunks, and builds a RAG index of ~300k documents. Pipeline uses critic/evaluator models, DPO alignment from ~1k human preferences and 5k QA annotations, and a fact‑check loop that reprompts to reduce hallucination (the rerun output wins ~70% of the time); total compute is ~30–45k GPU hours and a public pilot and full open‑source release are scheduled this autumn.
AI-written summary · about
chapters (20)
- 0:00EVE project introduction
- 1:26Earth observation importance
- 2:50Contributors and timeline
- 3:54Target audience
- 4:36Training pipeline
- 6:20Earth observation corpus
- 8:50Instruction datasets
- 10:13Fine-tuning domain LLM
- 11:17Mistral 3.2 24B choice
- 12:20Synthetic instruction data
- 14:27Synthetic data pipeline
- 16:39Summarization training
- 17:44Preference alignment
- 18:47Fine-tuning experiments
- 20:56RAG database connection
- 21:39Fact-checking module
- 23:24Tool usage retrieval
- 26:57Evaluation benchmarks
- 28:43Project overview
- 30:27Q&A
related talks
29mLeveraging open-source LLMs for production
21mDemocratizing AI with Open Source Multi-Agent Systems
40mDeploying Your Own Language Models: A Use Case Demo with Open Source Tools
25mSet Up an Open Source RAG in 30 Minutes
34mMapping the Open Model Landscape in 2025
15mOpenAssistant's Opensource LLMs Saga
13mHow Open-Source vLLM Topped the Artificial Analysis Leaderboard
22mBuild an Agent with Eve: The Open Agent Framework