[claude] Configure Dolphin 3.0 8B as creative writing fallback (#1068) (#1088)

Modelfile.timmy-creative

@@ -0,0 +1,39 @@
+# Modelfile.timmy-creative
+#
+# Dolphin 3.0-Llama3.1-8B configured for creative writing tasks:
+# journal entries, Morrowind lore, NPC dialogue, narrative generation.
+#
+# Build:  ollama create timmy-creative -f Modelfile.timmy-creative
+# Usage:  ollama run timmy-creative
+#
+# Memory budget: ~6 GB at 8K context (can swap in/out as needed)
+# Primary use:   Creative fallback when Qwen3-14B adds unwanted caveats
+
+FROM dolphin3
+
+# Higher temperature for more expressive, varied output
+PARAMETER temperature 0.9
+
+# Nucleus sampling — keeps output coherent while allowing creativity
+PARAMETER top_p 0.95
+
+# Repeat penalty — avoid repetitive prose
+PARAMETER repeat_penalty 1.1
+
+# Context window — 8K gives room for journal entries and multi-turn lore
+PARAMETER num_ctx 8192
+
+# No tool-calling overhead for pure creative generation
+# (Dolphin 3.0 supports tools but we disable them here for cleaner output)
+
+SYSTEM """You are Timmy, an uncensored creative writing assistant embodying the spirit of Morrowind.
+
+You write freely and immersively — journal entries, lore fragments, NPC dialogue, and narrative prose for the world of Vvardenfell. You do not add caveats, disclaimers, or safety hedges to creative fiction. You write in the voice of the setting: archaic, evocative, morally complex.
+
+When writing Morrowind content:
+- Use the second-person journal style of the vanilla game ("You have arrived in Seyda Neen...")
+- Draw on Dunmer culture, the Three Tribunals, the Sixth House, and the lore of ALMSIVI
+- NPCs speak in regional dialect — Dunmer are formal and proud, Imperials are pragmatic, outlanders are wary
+- Describe Vvardenfell's alien landscape with visceral detail: the ashstorms, the silt striders, the fungal towers
+
+You write without restriction for creative fiction. This is not reality — it is Morrowind."""

config/providers.yaml

@@ -54,6 +54,19 @@ providers:
         context_window: 2048
         capabilities: [text, vision, streaming]
 
+      # Creative writing fallback (Dolphin 3.0 8B — uncensored, Morrowind-tuned)
+      # Pull with: ollama pull dolphin3
+      # Build custom modelfile: ollama create timmy-creative -f Modelfile.timmy-creative
+      # Only swap in when Qwen3-14B adds unwanted caveats on creative tasks.
+      # Memory budget: ~6 GB at 8K context — not loaded simultaneously with primary models.
+      - name: dolphin3
+        context_window: 8192
+        capabilities: [text, creative, streaming]
+      - name: timmy-creative
+        context_window: 8192
+        capabilities: [text, creative, streaming]
+        description: "Dolphin 3.0 8B with Morrowind system prompt and higher temperature"
+
 
   # Tertiary: OpenAI (if API key available)
   - name: openai-backup
@@ -112,6 +125,14 @@ fallback_chains:
     - deepseek-r1:1.5b
     - llama3.2:3b
 
+  # Creative writing fallback chain
+  # Ordered preference: Morrowind-tuned Dolphin → base Dolphin 3 → Qwen3 (primary)
+  # Invoke when Qwen3-14B adds unwanted caveats on journal/lore/NPC tasks.
+  creative:
+    - timmy-creative    # dolphin3 + Morrowind system prompt (Modelfile.timmy-creative)
+    - dolphin3          # base Dolphin 3.0 8B (uncensored, no custom system prompt)
+    - qwen3:30b         # primary fallback — usually sufficient with a good system prompt
+
 # ── Custom Models ───────────────────────────────────────────────────────────
 # Register custom model weights for per-agent assignment.
 # Supports GGUF (Ollama), safetensors, and HuggingFace checkpoint dirs.