---

base_model: HuggingFaceTB/SmolLM2-360M-Instruct

license: apache-2.0

tags:

• gguf
• marola
• ocean

pipeline_tag: text-generation

---

marola-sea-tiny-GGUF

A GGUF release of a marola-sea checkpoint — a small model tuned on marola's own

question/answer shape (open-water swim conditions, safety, sea life), served locally through

Ollama alongside marola's own RAG corpus and deterministic scoring

(Recommender/Swimability, never the model itself).

Honest framing (MIP-0025 §6, §8): tuning changes tone and format reliability, not factual

grounding. This model does not replace marola's RAG corpus (knowledge/, cited answers only) or

its Reviewer pass, and it is not a standalone safety authority — treat any first-aid or

hazard answer as a starting point, not a substitute for a lifeguard or emergency services (marola's

own answers carry this caveat automatically via the MIP-0022 safety footer; this raw checkpoint,

used outside marola, does not).

Files

• marola-sea-tiny-Q8_0.gguf (386 MB)

CHECKSUMS (sha256) ships alongside these files — pin a specific hash in your own Modelfile

rather than a bare filename, so re-quantizing upstream can't silently change what you run.

Use with Ollama

ollama run hf.co/h0ffmann/marola-sea-tiny-GGUF
# or a specific quant tag, e.g.:
ollama run hf.co/h0ffmann/marola-sea-tiny-GGUF:Q4_K_M

Base model and training

Fine-tuned from HuggingFaceTB/SmolLM2-360M-Instruct via LoRA (finetune/train_lora.py

in h0ffmann/marola) on a small, hand-built dataset derived

from marola's own DSPy-compiled demos, sea-lore entries, and knowledge-corpus Q&A

(finetune/build_dataset.py) — a few dozen examples, enough to teach format and tone, not facts.

No just benchmark numbers recorded yet for this checkpoint — see docs/benchmarks/ in the source repo before trusting this over a plain base model.

Licence

Base model licence: apache-2.0. See the base model's own repo for the full licence text.