Coralfil

Domain-specific AI for the ocean. We train and serve open-weights LLMs on marine science, fisheries, and maritime regulatory compliance — with citations, provenance audits, and benchmarks that disclose their own weaknesses.

Models (Apache 2.0)

Model	Base	What it's for	Key verified results
Atlantis-Pyramid-v2	Qwen2.5-32B-Instruct + DeepSeek-R1-Distill-Qwen-32B	128k-ctx marine-science foundation model	IFEval prompt 55.0/64.0 (n=100, lm-eval 0.4.13) · AboveBoard-Marine-52 97.1%/100% MC (internal 2026-10-11)
Atlantis-Obelisk-14B	unsloth/Qwen2.5-14B-Instruct-bnb-4bit	compliance & action-synthesis coprocessor	IFEval prompt 52.31/55.27 (full 541 set) · AboveBoard MC 22/22*

* position-bias caveat disclosed on the benchmark card; v2 key rebalance in progress.

Datasets & benchmarks

AboveBoard-Marine-52 — 52 auditable marine-safety/regulatory scenarios with rubrics
Atlantis Ocean Reasoning SFT & DPO — training corpus (CC-BY-4.0, provenance-cleared)
Pacific Marine Taxonomy & CCME Water Quality — 64 species w/ YOLO labels + WoRMS AphiaIDs; 953 CCME records

Principles

Open weights (Apache 2.0), open data where licence allows (CC-BY-4.0 / OGL-Canada)
Every published number measured by us, with protocol and n disclosed
Correction logs on our cards when we get something wrong

Serving: os.coralfil.com and atlantis-llm.io· Contact: roland.poulin@coralfil.com