NatureCode Ocean Life

A 9.91 billion parameter foundation model specialized for marine biodiversity, ocean ecosystems, and aquatic life sciences.

Model Description

NatureCode Ocean Life is a decoder-only transformer model trained on curated marine science data from authoritative biodiversity databases. It understands species taxonomy, marine habitats, ecological relationships, and oceanographic concepts.

Model Details

Model Type: Decoder-only Transformer (Causal LM)
Parameters: 9.91B
Context Length: 4,096 tokens
Training Data: 871,304 examples
Language: English
License: Apache 2.0

Training Data

Included Sources

Source	Description	Size
OBIS	Ocean Biodiversity Information System - species occurrence records	62GB
WoRMS	World Register of Marine Species - taxonomic database	4K records
FishBase	Comprehensive fish species database	4K records
GBIF	Global Biodiversity Information Facility (marine subset)	1.8K records
arXiv	Marine science research papers	~1,000 papers
OceanInstruct	Marine Q&A instruction data	4.9MB
Text Corpus	Curated marine biology documents	3.4GB

Data Gaps (Not Yet Included)

The following data types are planned for future versions:

Soil & Terrestrial Ecology - Land ecosystems, soil microbiomes, forest biodiversity
Freshwater Systems - Rivers, lakes, wetlands, freshwater fish
Climate Data - Ocean-atmosphere interactions, temperature records, climate projections
Satellite Imagery - Ocean color, chlorophyll, sea surface temperature
Acoustic Data - Marine mammal vocalizations, underwater soundscapes, bioacoustics
Chemical Oceanography - Nutrient cycles, pH measurements, dissolved oxygen
Genetic/eDNA - Environmental DNA, genomic sequences, phylogenetic data
Historical Records - Historical fisheries data, expedition logs, museum collections
Indigenous Knowledge - Traditional ecological knowledge, local species names

Intended Uses

Primary Use Cases

Species identification and taxonomy queries
Marine habitat and distribution research
Biodiversity assessment support
Ocean ecosystem understanding
Marine conservation planning
Fisheries science applications
Educational content about marine life

Out of Scope

Medical or dietary advice about seafood
Real-time species identification from images (text-only model)
Legal or regulatory compliance decisions
Replacement for authoritative taxonomic databases

Training Procedure

Hardware

8x NVIDIA H100 80GB GPUs
Distributed Data Parallel (DDP) training

Hyperparameters

Batch Size: 16 per device
Gradient Accumulation: 4 steps
Effective Batch Size: 512
Learning Rate: 3e-4 (with warmup)
Warmup Steps: 1,000
Total Steps: 50,000
Precision: BF16 mixed precision
Optimizer: AdamW

Evaluation

Model evaluation on marine biology benchmarks is ongoing. Performance metrics will be updated as they become available.

Limitations and Biases

Known Limitations

Marine Focus: Limited knowledge of terrestrial and freshwater ecosystems
Language: Training data primarily in English
Temporal: May not include species discovered after training data cutoff
Geographic: Data coverage may be uneven across ocean regions
Depth: Deep-sea species may be underrepresented due to limited research data

Ethical Considerations

Should not be used as the sole basis for conservation decisions
Predictions should be validated against authoritative databases (WoRMS, OBIS, FishBase)
May perpetuate biases present in historical biodiversity records

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("naturecode/ocean-life")
tokenizer = AutoTokenizer.from_pretrained("naturecode/ocean-life")

prompt = "What are the main threats to coral reef ecosystems?"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Citation

@misc{naturecode-ocean-life-2026,
  title={NatureCode Ocean Life: A Foundation Model for Marine Biodiversity},
  author={NatureCode Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/naturecode/ocean-life}
}

Acknowledgments

Training data sourced from:

Contact

For questions or feedback, please open an issue on the model repository.

Downloads last month: -; Downloads are not tracked for this model. How to track