aidnd-location

aidnd-location

QLoRA adapter for Qwen3-14B, merged and exported to GGUF (Q4_K_M), for the AI-DnD Engine. It is the location_writer role: it turns a structured location spec into a terse, grounded prose description for a Russian D&D frontier setting — including a one-level tree of sub-locations (a tavern's common room / kitchen / cellar, a ruin's cellar pit).

• Base: Qwen3-14B · Method: QLoRA (4-bit), merged → GGUF Q4_K_M
• Input: location parameters — type, size, material, condition, smell, sound, light, air, mood, affordances, … (rendered by location_user)
• Output: 3–5 sentences, concrete sensory detail, no purple prose, no invented elements, no numbers — живой, простой, заземлённый стиль
• Dataset: 534 nodes (154 locations + 380 sub-locations) across 20+ archetypes, parametrized as a tree (datasets/location/)

Use with Ollama

huggingface-cli download Illaitar/aidnd-location location-q4_k_m.gguf Modelfile --local-dir aidnd-location
cd aidnd-location && ollama create aidnd-location -f Modelfile
ollama run aidnd-location "Локация «Кузница» (тип: smithy; запах: угар, окалина; свет: багровый от горна). Опиши место для рассказчика: 3-5 предложений."

Trained with the QLoRA pipeline in training/ (Unsloth → merge 16-bit → convert_hf_to_gguf → quantize → Ollama). One of several per-role adapters; see the model overview.