The project bundles a domain-adapted Mistral Small 3.2 model, curated corpora, MCQA and open-ended QA benchmarks, RAG integration, and a hallucination-detection pipeline.
AI Quick Take
- EVE-Instruct is a 24B model adapted from Mistral Small 3.
- The release includes datasets, domain benchmarks, RAG and hallucination-detection tooling, plus an API/GUI deployment used by ~350 pilot users and planned open-source publication.
EVE (Earth Virtual Expert) published EVE-Instruct, a 24B domain-adapted LLM built on Mistral Small 3.2 and tailored for Earth Observation and Earth Sciences question answering and reasoning. The project pairs the model with curated training corpora, new domain-specific benchmarks, and deployment tooling.
The authors report that EVE-Instruct outperforms comparable models on their newly constructed MCQA, open-ended QA, and factuality benchmarks while maintaining general capabilities. The release also integrates retrieval-augmented generation and a hallucination-detection pipeline, and the system is available via an API and GUI used by roughly 350 pilot users.
All models, datasets, and code are slated for open release under permissive licenses on Hugging Face and GitHub. Readers should inspect the published benchmarks and await independent evaluations to confirm the paper's performance claims and to assess the model's suitability for operational Earth-intelligence tasks.