POIEO-GGUF

POIEO is a lightweight bilingual creative-writing assistant, continually pre-trained and fine-tuned from Liquid AI's LFM2.5-2.6B-Base. It's built to help writers overcome creative blocks, develop ideas, and continue stories naturally in English and Yoruba — including code-switched writing that shifts between the two mid-scene.

POIEO is optimized for local, on-device inference on consumer hardware, not cloud deployment. Development and testing targeted a mid-range laptop class (Intel Core i5, 8GB RAM, integrated graphics, no discrete GPU) to keep the bar realistic for the hardware most writers actually have.

✨ What POIEO can do

• ✍️ Generate and develop creative ideas
• 📖 Continue stories from where the writer stopped
• 📝 Write creatively in English and Yoruba, including within a single scene
• 🔄 Maintain context across a conversation
• 💻 Run locally on consumer hardware, entirely offline
• 🔒 Keep creative work on-device — nothing sent to a remote service

How POIEO was built

Training was continued pre-training (CPT) followed by supervised fine-tuning (SFT) — there is no DPO or other preference-tuning stage in this release. That's worth knowing going in: the model hasn't been optimized against pairwise preference judgments, so tone, verbosity, and consistency are shaped entirely by the SFT data itself rather than a separate alignment pass. If you notice uneven quality across similar prompts, this is the likely reason, and it's the natural next step for a future version.

1. CPT — LFM2.5-2.6B-Base was continually pre-trained on Yoruba text (quality-filtered Yoruba Wikipedia plus additional Yoruba corpora) to strengthen the base model's fluency and diacritic accuracy in Yoruba before any instruction tuning.
2. SFT — the CPT checkpoint was then fine-tuned on bilingual creative-writing prompt/response pairs (Yoruba, English, and bilingual code-switched examples), trained with LoRA via Unsloth, with loss computed only on the assistant's response.

Usage with llama.cpp

llama-cli -hf Obotu/POIEO-GGUF --jinja

Or download a specific .gguf file and run it directly. Recommended flags for constrained hardware (tune --threads to your actual physical core count):

llama-cli -m lfm2.5-yoruba-sft-final.Q5_K_M.gguf \
  --threads 4 --ctx-size 2048 \
  --batch-size 128 --ubatch-size 128 \
  --n-gpu-layers 999

--n-gpu-layers 999 offloads as many layers as fit to an Intel integrated GPU via llama.cpp's Vulkan or SYCL backend, if built with one enabled — this reduces CPU load and thermal pressure on laptops with Iris Xe/UHD graphics. On CPU-only builds, drop that flag.

Example

User:
Kọ ìtàn kúkúrú kan nípa ọ̀dọ́kùnrin kan tí ó rí àṣírí kan
ní abẹ́ ilé àgbà rẹ̀.

POIEO:
[Generates a Yoruba creative story...]

User:
Continue this story, keeping the narrator's voice in Yoruba and
switching to English for the foreign character's dialogue:

Adé wo ọ̀nà òkùnkùn tó wà níwájú rẹ̀...

Limitations

POIEO is an experimental fine-tuned model. Known limitations:

• No DPO/preference-tuning stage — outputs can be inconsistent in tone or quality across similar prompts (see "How POIEO was built" above)
• Identity/self-description questions ("who are you?") are not yet reliably handled — the model may give a vague or off-topic answer rather than a consistent persona response
• Can produce factual inaccuracies — this is a creative-writing model, not a factual-knowledge model
• Yoruba grammar and diacritics are generally strong but not guaranteed correct in every output
• May lose coherence in very long conversations
• Output quality varies with sampling parameters (temperature, top-p) — the examples above used moderate settings; adjust for your use case

POIEO should be treated as a creative writing companion, not a factual authority or a finished product — this is an active work in progress.

Roadmap

• DPO/preference-tuning pass to improve consistency and reduce the identity-question gap above
• Expand beyond Yoruba and English to additional African languages
• Explore multimodal and voice-based creative workflows

Credits

Built on Liquid AI's LFM2.5-2.6B-Base and fine-tuned with Unsloth.