Cookbooks // Machine Learning & Training

Marine ML Training & Sensor Data Curation

Training reliable language models for oceanography requires clean training corpora grounded in real physical observations. These recipes demonstrate automated sensor data quality control, license validation, and parameter-efficient 4-bit LoRA fine-tuning for marine domains.

GitHub Source & Cookbooks Directory

Runnable Python scripts and schema files are located in the public repository:

cookbooks/ml-training/
curate_marine_dataset.py • train_marine_lora.py
Browse ml-training on GitHub →

Recipe 1: Marine Data Curation

Ocean telemetry from surface buoys and moorings often contains uncalibrated electrical spikes or biofouling drift. The curate_marine_dataset.py module enforces physical oceanographic boundaries, confirms permissible open licenses (such as CC-BY-4.0 and OGL Canada), and exports JSONL instruction pairs for fine-tuning.

Physical Validation Limits

Sensor readings outside realistic seawater thresholds are automatically flagged and clamped:

  • Temperature: -2.0 °C to 36.0 °C (liquid seawater freeze point to extreme surface heating).
  • Practical Salinity: 0.0 to 42.0 PSU (estuarine fresh runoff to hypersaline basins).
  • pH: 6.8 to 8.6 (acidified sediment porewater to surface algal bloom baselines).
  • Dissolved Oxygen: 0.0 to 18.0 mg/L (anoxic dead zones to supersaturated breaking surf).

Python Execution Example

from curate_marine_dataset import MarineDataCurator

curator = MarineDataCurator(output_dir="./curated_dataset")

# Verify license and institution metadata
provenance = curator.validate_provenance({
    "source": "Pacific Coastal Observation Buoy",
    "license": "CC-BY-4.0",
    "url": "https://example.org/station-data",
    "platform": "subsurface_mooring",
    "institution": "Salish Sea Marine Institute",
    "date": "2026-11-14",
})

# Sanitize raw observation packet
raw_reading = {
    "sensor_id": "STATION-42",
    "timestamp": "2026-11-14T12:00:00Z",
    "latitude": 49.31,
    "longitude": -123.18,
    "depth_meters": 18.5,
    "temperature_c": 10.8,
    "salinity_psu": 29.4,
    "ph_level": 7.91,
    "dissolved_oxygen_mgl": 6.4,
    "turbidity_ntu": 1.2,
}

sanitized_record = curator.sanitize_record(raw_reading, provenance)
curator.export_to_jsonl([sanitized_record], "training_samples.jsonl")

Recipe 2: Marine LLM 4-Bit LoRA Training

Fine-tuning large language models (7B to 32B parameters) directly in 16-bit precision requires expensive high-VRAM clusters. The train_marine_lora.py script configures a parameter-efficient QLoRA pipeline using 4-bit NormalFloat (NF4) quantization and double quantization.

Memory Footprint Comparison

Base weights are compressed into 4-bit representations, while low-rank adapter matrices are injected into attention and MLP projections:

16-Bit Full Fine-Tuning
~14.5 GB VRAM (Weights)

Requires 24 GB+ enterprise GPU for gradient states and activations.

4-Bit QLoRA Pipeline
~3.5 GB VRAM (Weights)

Trainable on local 8 GB to 16 GB workstations with under 45M trained parameters.

QLoRA Hyperparameter Specification

from train_marine_lora import MarineLoraPipeline, MarineTrainingConfig, MarineLoraConfig

training_config = MarineTrainingConfig(
    model_name_or_path="Qwen/Qwen2.5-7B-Instruct",
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-4,
    num_train_epochs=3,
    output_dir="./marine_lora_checkpoint",
)

lora_config = MarineLoraConfig(
    r=16,
    lora_alpha=32,
    lora_dropout=0.05,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
)

pipeline = MarineLoraPipeline(training_config=training_config, lora_config=lora_config)
pipeline.verify_and_export_manifest("./pipeline_manifest.json")

Related Resources