---
language:
- en
license: apache-2.0
library_name: transformers
tags:
- ocean
- ocean-restoration
- ocean-climate
- aquaculture
- water-management
- ocean-acidification
- marine-biology
- coralfill
- qwen
- atlantis
- lora
- sft
base_model: Qwen/Qwen2.5-14B-Instruct
pipeline_tag: text-generation
inference: false
---

# Atlantis-Pyramid 14B (Glyph)

**Atlantis-Pyramid 14B (Glyph)** is the flagship model of the Coralfil OS Atlantis project.
It is a fine-tuned version of `Qwen/Qwen2.5-14B-Instruct` (~14.7B parameters, verified 2026-08-10: 48 layers, hidden 5120),
specialized for marine restoration science, aquaculture chemistry,
dynamic oceanography modeling, and coral reef ecosystem analysis.

## Model Details

| Property | Value |
|---|---|
| **Base Model** | `Qwen/Qwen2.5-14B-Instruct` |
| **Parameters** | ~14.7B (renamed from '32B' — measured weight dims are 14B-class) |
| **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 |

## Training

The model was fine-tuned using Supervised Fine-Tuning (SFT) with the `trl`
library's `SFTTrainer`. LoRA adapters were trained in 4-bit precision 
for memory efficiency and then merged into the base model for inference.

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

## Usage

### With Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "Coralfil/Atlantis-Pyramid-32B",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Coralfil/Atlantis-Pyramid-32B")

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

```bash
vllm serve Coralfil/Atlantis-Pyramid-32B --dtype bfloat16 --max-model-len 4096
```

## Intended Use

This model is designed for research and operational use within the Coralfil OS
ecosystem. It excels at:

- Marine biology question answering
- Coral reef ecosystem analysis
- Aquaculture parameter optimization
- Oceanographic data interpretation
- Environmental compliance assessment

## Limitations

- This model is specialized for marine science domains and may underperform
  on unrelated tasks compared to the general-purpose base model.
- Outputs should be validated by domain experts for critical decisions.
- Not intended for medical or safety-critical applications.

## License

Apache 2.0 -- see [LICENSE](LICENSE) for details.

## Citation

```bibtex
@misc{atlantis-pyramid-2026,
  title={Atlantis-Pyramid-32B: A Fine-Tuned 32B Model for Marine Science},
  author={Coralfil Team},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/Coralfil/Atlantis-Pyramid-32B}
}
```