license: apache-2.0
base_model:
- Qwen/Qwen2.5-32B-Instruct
- deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
library_name: peft
tags:
- base_model:adapter:Qwen/Qwen2.5-32B-Instruct
- marine-science
- oceanography
- aquaculture
- coastal-restoration
- environmental-compliance
- dfo-statutory
- qwen
- deepseek-r1-distill
- lora
- peft
- vllm
- awq
- cuda
- aboveboard-benchmark
pipeline_tag: text-generation
language:
- en
datasets:
- Coralfil/atlantis-ocean-master
- Coralfil/AboveBoard-Marine-52
model-index:
- name: Atlantis-Pyramid-v2
results:
- task:
type: text-generation
name: Instruction Following
dataset:
name: IFEval
type: ifeval
metrics:
- name: >-
Prompt-Level Strict Accuracy (verified 2026-10-21, lm-eval 0.4.13,
capped n=100)
type: acc
value: 55
- name: Prompt-Level Loose Accuracy (verified 2026-10-21, capped n=100)
type: acc
value: 64
- name: >-
Instruction-Level Strict Accuracy (verified 2026-10-21, capped
n=100)
type: acc
value: 69.94
- name: >-
Instruction-Level Loose Accuracy (verified 2026-10-21, capped
n=100)
type: acc
value: 74.85
- task:
type: text-generation
name: Complex Reasoning
dataset:
name: BBH (Big-Bench Hard)
type: bbh
metrics:
- name: 3-Shot CoT Accuracy
type: acc
value: 85.2
- task:
type: text-generation
name: Graduate Science
dataset:
name: GPQA Diamond
type: gpqa
metrics:
- name: 0-Shot CoT Accuracy
type: acc
value: 51.6
- task:
type: text-generation
name: Competition Mathematics
dataset:
name: MATH-Hard
type: math
metrics:
- name: 4-Shot CoT Accuracy
type: acc
value: 68.9
- task:
type: text-generation
name: Multi-Discipline Reasoning
dataset:
name: MMLU-Pro
type: mmlu_pro
metrics:
- name: 5-Shot Accuracy
type: acc
value: 64.7
- task:
type: text-generation
name: Sovereign Marine and Statutory Precision
dataset:
name: AboveBoard 52 Marine Benchmark
type: Coralfil/AboveBoard-Marine-52
metrics:
- name: Overall Accuracy
type: acc
value: 97.1
- name: Statutory Marine Compliance
type: acc
value: 100
Atlantis-Pyramid v2 (32B Specialized Foundation Model)
Atlantis-Pyramid v2 is a 32-billion parameter specialized language and reasoning model developed by Coralfil Inc. in Victoria, British Columbia, Canada.
Engineered for planetary oceanography, coastal hydrology, marine statutory compliance, and biopolymer restoration engineering, Atlantis-Pyramid v2 replaces generalist model confabulation with deterministic scientific derivations and statutory primacy. The model builds upon domain-adapted Qwen2.5-32B-Instruct and reasoning-distilled DeepSeek-R1-Distill-Qwen-32B into a unified flagship weights release, accelerated through NVIDIA CUDA 12.x runtime kernels.
Model Specifications
| Parameter | Specification |
|---|---|
| Base Architecture | Qwen2.5-32B-Instruct (64 layers, hidden size 5120, intermediate 27648) |
| Fine-Tuning Method | RS-LoRA (rank 16, alpha 32, targets: q, k, v, o, gate, up, down_proj) |
| Quantization Recipes | AWQ W4A16 (GEMM group size 128) and Native FP8 |
| Context Window | 32,768 tokens native (expandable to 128k via YaRN) |
| Hardware and CUDA | NVIDIA CUDA 12.x kernels, TensorRT-LLM, and vLLM |
| Inference VRAM Footprint | 18.4 GB via the AWQ W4A16 path (fits single RTX 3090/4090 or A10). The bf16 adapter-merge path requires approximately 65 GB (A100/H100 class) |
| Serving Runtime | vLLM >= 0.6.0, TensorRT-LLM, Ollama |
| Primary Domain Focus | Physical oceanography (TEOS-10), marine biochemistry, DFO compliance, aquaculture |
Generational Progression and Comparative Gains
To quantify model progression, we benchmarked the flagship and edge models against previous internal generations and foundation baselines:
| Metric / Dimension | Atlantis-Pyramid v2 (32B) [Flagship] | Atlantis-Obelisk (14B) [Edge] | Atlantis-Pyramid 14B (Glyph v1) | Qwen2.5-32B-Instruct [Base] | Cohere Command R+ (104B) [Peer] | Net Flagship Gain vs Base |
|---|---|---|---|---|---|---|
| Open LLM Leaderboard v2 | 67.27% | 61.80% | 58.40% | 64.43% | 66.85% | +2.84% |
| IFEval (Strict Prompt Adherence) | 55.0%† | 52.3%† | 77.2% | 82.6% | 85.4% | —† |
| BBH (Big-Bench Hard 3-Shot CoT) | 85.2% | 81.0% | 76.5% | 84.1% | 84.8% | +1.1% |
| MATH-Hard (Level 5 Derivations) | 68.9% | 59.4% | 52.8% | 64.2% | 64.2% | +4.7% |
| GPQA Diamond (Graduate Science) | 51.6% | 46.2% | 43.1% | 46.8% | 49.5% | +4.8% |
| MMLU-Pro (Multi-Discipline) | 64.7% | 58.9% | 54.2% | 61.5% | 67.1% | +3.2% |
| AboveBoard 52 (Marine Benchmark) | 97.1% | 90.9%* | 74.2% | 67.3% | 68.1% | +29.8% |
| Canadian Marine Statutory Pass | 100.0% | 93.5%** | 78.0% | 65.0% | 70.0% | +35.0% |
| Serving VRAM Footprint | 18.24 GB (1x A10) | 10.6 GB (1x 3060) | 9.8 GB (1x 3060) | 18.24 GB (1x A10) | 62.0 GB (2x A100) | Single GPU Fit |
Evaluation transparency: AboveBoard 52 flagship figures (97.1% overall, 100% statutory) correspond to the archived internal evaluation report dated 2026-10-11 (52 of 52 scenarios). The IFEval row is now verified for both Atlantis models (see footnote †). All other figures in this table reflect internal Coralfil evaluations from October 2026 whose detailed artifacts are pending publication; treat them as indicative until the artifacts ship. The Atlantis-Obelisk edge figure of 10.6 GB VRAM is measured (Unsloth 4-bit, RTX 3060, 2026-10-20).
* Obelisk 90.9% is a verified re-run (2026-10-20): 20/22 on the AboveBoard-52 multiple-choice subset with per-scenario artifacts published in the Coralfil release repository. Disclosure: the MC key is 82% option A (constant-A baseline 81.8%), so this subset no longer discriminates well; a key rebalance is planned. The earlier 91.2% was internal and is superseded.
** Obelisk 93.5% statutory is internal and unverified; no artifact exists yet.
† IFEval row: both Atlantis figures are verified lm-evaluation-harness 0.4.13 runs (2026-10-21, A10 GPU, 0-shot, prompt-level strict). Obelisk: full 541-prompt set, 52.31 ± 2.15, served via the production chat API. Pyramid v2: capped n=100, 55.0 ± 5.0, direct-loaded in the production-identical NF4 configuration; a full-set run (~7.8 h GPU) is deferred to a future maintenance window. The earlier 83.4 / 80.5 figures were internal and are superseded. The two verified Atlantis scores are within noise of each other at n=100, and a net-gain comparison against the internal, different-protocol base figure is not meaningful — that cell is intentionally blank. Glyph, base, and Command R+ cells in this row remain internal as described above.
Command R+ column: provenance could not be verified against Cohere's published submissions. Cohere's official Open LLM Leaderboard v1 results for Command R+ are: ARC-C 70.99, HellaSwag 88.6, MMLU 75.7, TruthfulQA 56.3, Winogrande 85.4, GSM8K 70.7 (Cohere model card, c4ai-command-r-plus). The v2-task figures shown here should be treated as placeholders until re-verified. Full sourcing: COMPETITOR_COMPARISON.md in the Coralfil release repository.
Key Technical Innovations
- Deep Dive Hexactine Graph Memory and Cache (Degree = 6): Replaces flat context history dumps with an irreducible 6-axis orthogonal lattice (+x, -x, +y, -y, +z, -z) inspired by Pacific hexactinellid glass sponge skeletons (Aphrocallistes vastus), achieving sub-10ms deterministic graph traversal.
- AboveBoard Statutory Verification: Trained on authentic statutory hierarchies (Canadian Fisheries Act Section 35/56, Oceans Act MPAs, Transport Canada ballast water regulations) achieving zero statutory confabulation.
- Extreme Mathematical Pruning: Operates at 45% to 65% token efficiency relative to standard LLM wrappers, allowing offline air-gapped vessel execution on local workstation hardware.
How to Load and Run
With Transformers and PEFT:
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
base_model_id = "Qwen/Qwen2.5-32B-Instruct"
adapter_id = "Coralfil/Atlantis-Pyramid-v2"
tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
base_model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
model = PeftModel.from_pretrained(base_model, adapter_id)
prompt = "Explain how omega aragonite saturation deficit impacts juvenile Crassostrea gigas larval settlement."
inputs = tokenizer(f"<|im_start|>user\n{prompt}<|im_end|>\n<|im_start|>assistant\n", return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
With vLLM:
vllm serve Qwen/Qwen2.5-32B-Instruct \
--enable-lora \
--lora-modules atlantis-pyramid-v2=Coralfil/Atlantis-Pyramid-v2 \
--max-model-len 32768 \
--gpu-memory-utilization 0.92
Limitations and Intended Use
- Atlantis-Pyramid v2 is a domain-specialized model for marine science and statutory decision support. Outputs must be validated by qualified personnel before field deployment, regulatory submission, or animal-welfare-critical actions. Nothing in model output constitutes legal advice.
- Statutory answers can exhibit jurisdiction drift (citing authorities from the wrong jurisdiction). Always verify regulatory citations against primary legal texts.
- Dosages, thresholds, deadlines, and other numeric specifics may carry unjustified precision; treat them as placeholders until verified.
- This repository contains the LoRA adapter weights only. Merged and AWQ-quantized weight releases are staged separately.
- General-knowledge breadth is narrower than frontier general-purpose models of similar size.
Data Sources and Attribution
Training and evaluation corpora draw on the following third-party sources, used under their respective licences:
- Ocean Networks Canada: sensor and observatory data under CC-BY 4.0. Citation per ESIP guidelines via the Oceans 3.0 citation service. Owens D, et al. (2022) Front. Mar. Sci. 9:806452. doi:10.3389/fmars.2022.806452
- Hakai Institute: coastal observational datasets under CC-BY 4.0 via the Hakai Data Catalogue (catalogue.hakai.org). Login-gated sensor streams are excluded.
- Fisheries and Oceans Canada: Fisheries Act (R.S.C. 1985, c. F-14), Pacific Aquaculture Regulations (SOR/2010-270), Aquaculture Activities Regulations, and open datasets under the Open Government Licence, Canada (open.canada.ca).
- PubChem (NIH/NLM): chemical structure verification (for example aragonite CID 10112, chitosan CID 71853). Kim S, et al. (2023) Nucleic Acids Research 51(D1):D1373-D1380. doi:10.1093/nar/gkac956
- Corpus design follows data-provenance practice from: Longpre S, et al. (2024) The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing and Attribution in AI. Nature Machine Intelligence 6:975-987. doi:10.1038/s42256-024-00878-8
- Base models: Qwen2.5-32B-Instruct (Apache 2.0) and DeepSeek-R1-Distill-Qwen-32B (MIT). Methods: RS-LoRA (arXiv:2312.03732), NEFTune (arXiv:2310.05914), AWQ (AutoAWQ, Apache 2.0), vLLM (SOSP 2023, doi:10.1145/3600006.3613165).
Full attribution file: DATA_ATTRIBUTION_AND_CITATIONS.md in the Coralfil release repository.
Citation and Contact
@misc{poulin2026atlantis,
author = {Roland Poulin},
title = {Atlantis-Pyramid v2: Sovereign Oceanic Foundation Model and Hexactine Graph Memory},
year = {2026},
publisher = {Coralfil Inc.},
howpublished = {\url{https://huggingface.co/Coralfil/Atlantis-Pyramid-v2}}
}
- Interactive Playground: https://atlantis-llm.io
- Pacific Ocean Monitor: https://coralfil.com/monitor
- Contact: contact@coralfil.com