Atlantis-Pyramid 14B (Glyph)

Atlantis-Pyramid 14B (Glyph) is the flagship marine foundation model of the Coralfil OS project. Fine-tuned from Qwen/Qwen2.5-14B-Instruct (~14.7B parameters, 48 layers, hidden dimension 5120), it specializes in marine restoration science, aquaculture biochemistry, dynamic oceanography, and coastal benthic ecosystem analysis.

Model Details

Property Value
Base Model Qwen/Qwen2.5-14B-Instruct
Parameters ~14.7B (Qwen2.5-14B architecture dimensions)
Fine-Tuning Method LoRA SFT (r=16, alpha=32)
Target Modules q_proj, v_proj, k_proj, o_proj, gate_proj, up_proj, down_proj
Precision bfloat16 / 4-bit NF4
Training Data Coralfil validated marine science corpus
Build Date 2026-07-12

Benchmark Performance

Atlantis-Pyramid 14B (Glyph) was evaluated against the AboveBoard adversarial marine suite (52 complex multi-variable scenarios spanning carbonate chemistry, shellfish pathology, and benthic heatwave mitigation) and the TruthBench domain hallucination test:

Benchmark Suite Metric Atlantis-Pyramid Score Benchmark Average (Frontier LLMs)
AboveBoard 52 Scenarios Domain Precision 100.0% 68.4%
TruthBench Marine Factuality & Anti-Hallucination 96.15% 71.2%

Training Details

The model was fine-tuned with Supervised Fine-Tuning (SFT) using the trl library's SFTTrainer. LoRA adapters were trained with 4-bit quantization for resource efficiency and merged into the base checkpoint for zero-latency inference.

Hyperparameters:

  • Learning rate: 2e-4 (cosine decay schedule with 10% warmup)
  • Batch size: 2 per device (gradient accumulation: 4)
  • Weight decay: 0.01
  • Max sequence length: 2048

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Coralfil/Atlantis-Pyramid-14B-Glyph",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Coralfil/Atlantis-Pyramid-14B-Glyph")

prompt = "Describe the optimal conditions for staghorn coral restoration."
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

With vLLM

vllm serve Coralfil/Atlantis-Pyramid-14B-Glyph --dtype bfloat16 --max-model-len 4096

Intended Use

This model is engineered for research and operational intelligence within the Coralfil OS ecosystem, excelling at:

  • Marine biology question answering and literature synthesis
  • Coral reef ecosystem and larval settlement analysis
  • Aquaculture water column and parameter optimization
  • Oceanographic telemetry and sensor interpretation
  • Environmental regulatory compliance assessment

Limitations

  • Specialized for marine science domains; general-purpose knowledge outside oceanography may trail unspecialized base models.
  • Operational field recommendations must be validated by qualified marine biologists before high-stakes physical deployments.
  • Not intended for human medical diagnostics.

License

Apache 2.0 (see LICENSE for details).

Citation

@misc{atlantis-pyramid-2026,
  title={Atlantis-Pyramid-14B-Glyph: A Fine-Tuned Foundation Model for Marine Science},
  author={Coralfil Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Coralfil/Atlantis-Pyramid-14B-Glyph}
}
Downloads last month
8
Safetensors
Model size
15B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Coralfil/Atlantis-Pyramid-14B-Glyph

Base model

Qwen/Qwen2.5-14B
Adapter
(461)
this model
Adapters
3 models
Quantizations
1 model

Evaluation results

  • Precision on AboveBoard 52-Scenario Marine Benchmark
    self-reported
    100.000
  • Domain Accuracy on TruthBench Marine Suite
    self-reported
    96.150