Atlantis-Obelisk: Sovereign Edge Marine Coprocessor

Atlantis-Obelisk is Coralfil's sovereign edge foundation coprocessor fine-tuned for physical oceanography, mariculture formulation, benthic restoration, and Canadian maritime statutory compliance (DFO, Transport Canada, IMO).

Optimized for low-latency onboard maritime vessel deployments, autonomous sensor nodes, and edge workstations.

Architecture and Specifications

• Base Architecture: Qwen 2.5 14B Instruct (4-bit base)
• Fine-Tuning: Rank-Stabilized LoRA (r=16, alpha=32) covering all 7 attention projection layers (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj).
• Context Length: 32,768 tokens (up to 128k via YaRN scaling).
• Domain Benchmarks: AboveBoard-52 multiple-choice, fully re-run with per-scenario artifacts on 2026-10-20: v1.2 adapter 20/22 (90.9%), v1.0 adapter 22/22, base Qwen2.5-14B-Instruct 20/22. Disclosure: the MC answer key is 82% option A, so a constant-A baseline scores 81.8%; this subset no longer discriminates well and a v2 key rebalance is planned. The earlier 91.2% / 93.5% figures were internal and are superseded by these verified runs.
• Instruction Following: IFEval full 541-prompt set, verified 2026-10-21 (lm-evaluation-harness 0.4.13, 0-shot, served via the production chat API): prompt-level strict 52.31 (± 2.15), loose 55.27; instruction-level strict 62.83, loose 65.83. The earlier 80.5% figure was internal and is superseded by this verified run. Results JSON and per-sample outputs are published in the Coralfil release repository.
• VRAM Footprint: 10.6 GB measured peak via the Unsloth 4-bit path (verified 2026-10-20 on a single NVIDIA RTX 3060 12 GB). A full bf16 merge requires approximately 30 GB.
• Interactive Playground: https://atlantis-llm.io
• Pacific Ocean Monitor: https://coralfil.com/monitor

Download and Run Locally

With Hugging Face CLI:

hf download Coralfil/Atlantis-Obelisk --local-dir ./atlantis-obelisk

Recommended: Unsloth 4-bit (verified on RTX 3060 12 GB):

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Coralfil/Atlantis-Obelisk",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

messages = [{"role": "user", "content": "Explain how omega aragonite saturation deficit impacts juvenile Crassostrea gigas larval settlement."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to("cuda")
outputs = model.generate(**inputs, max_new_tokens=512, use_cache=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Measured on RTX 3060 12 GB (2026-10-20): 10.6 GB peak VRAM, about 9.4 tokens/second.

Alternative: Transformers and PEFT at bf16 (requires approximately 30 GB VRAM):

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_model_id = "Qwen/Qwen2.5-14B-Instruct"
adapter_id = "Coralfil/Atlantis-Obelisk"

tokenizer = AutoTokenizer.from_pretrained(base_model_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))

Note: a plain BitsAndBytesConfig 4-bit load of the base can fail validation on 12 GB cards because the base repo ships its own quantization config. Use the Unsloth path above, a GPU with 16 GB or more, or CPU offload.

Version History

• v1.2 (2026-10-20, current): corrective patch, round 2. Continued SFT of the v1.0 adapter on a 28-record corrective dataset (36 steps, RSLoRA r=16 alpha=32, constant LR 1e-4, seq 1536). Cleared all four v1.0 quality flags on a 6-probe adversarial battery: jurisdiction conflation (including a BC+WDFW trap), invented numeric precision, <CHITOSAN> placeholder leak, persona phrasing. Trade-off: AboveBoard-52 MC moved 22/22 to 20/22; one miss is an ambiguous benchmark item, one is a genuine forced-choice error. Patch data, training script, and probe transcripts are published in the Coralfil release repository for full reproducibility.
• v1.1 (2026-10-20, superseded): corrective patch round 1 (18 records); cleared 2 of 4 flags, failed the adversarial jurisdiction trap, not released.
• v1.0 (2026-10-20): initial public adapter.

Limitations and Intended Use

Atlantis-Obelisk is a research and decision-support adapter trained on a compact 360-record domain corpus plus a 28-record corrective patch. The v1.2 probe battery (2026-10-20, RTX 3060) verified the fixes below, with these residual risks users must account for:

• Statute-level citation precision: jurisdiction routing (which agency and country applies) tested clean, but exact statute and chapter numbers (for example RCW or WAC references) can still be wrong. Verify every legal citation against the primary text.
• Numeric estimates: dosing and speciation answers now prefer ranges and measure-dose-retest procedures over invented constants, but any specific number should still be treated as an estimate until measured.
• Fill-in-the-blank and open-ended benchmark items are rubric-scored and pending adjudication; MC results above are final.

Outputs must be validated by qualified personnel before field deployment, regulatory submission, or animal-welfare-critical actions. Nothing in model output constitutes legal advice.

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: 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. Kim S, et al. (2023) Nucleic Acids Research 51(D1):D1373-D1380. doi:10.1093/nar/gkac956
• Base model: Qwen2.5-14B-Instruct (Apache 2.0). Fine-tuning stack: Unsloth (Apache 2.0) with RS-LoRA (arXiv:2312.03732).

Full attribution file: DATA_ATTRIBUTION_AND_CITATIONS.md in the Coralfil release repository.

Citation and Attribution

@software{poulin_2026_atlantis_obelisk,
  author = {Roland Poulin},
  title = {Atlantis-Obelisk: Sovereign Edge Marine Coprocessor},
  publisher = {Coralfil Inc.},
  url = {https://huggingface.co/Coralfil/Atlantis-Obelisk},
  year = {2026}
}

Roland Poulin, Director and Founder, Coralfil Inc.