feat: update model references to openbmb/MiniCPM-V-4.6 in documentation

README.md

@@ -14,7 +14,7 @@ pinned: false
 [![Deploy to Hugging Face Space](https://github.com/ASH1998/Rune-Goblin/actions/workflows/deploy-hf-space.yml/badge.svg?branch=master)](https://github.com/ASH1998/Rune-Goblin/actions/workflows/deploy-hf-space.yml)
 
 A tiny dungeon crawler where players draw spells in an invented symbolic
-language (**RuneLang**) and a fine-tuned [`openbmb/MiniCPM5-1B-SFT`](https://huggingface.co/openbmb/MiniCPM5-1B-SFT)
+language (**RuneLang**) and a fine-tuned [`openbmb/MiniCPM-V-4.6`](https://huggingface.co/openbmb/MiniCPM-V-4.6)
 acts as the **spell engine** — reading glyph combinations and emitting JSON
 that drives attacks, curses and game-state changes. Runtime visuals are not
 image-generated; spell metadata recolors, resizes, retargets and animates
@@ -27,7 +27,7 @@ See [`rune_goblin_plan.md`](./rune_goblin_plan.md) for the full design doc.
 ```
 Rune buttons (Gradio / React)
    → serialized rune sequence + game state
-   → fine-tuned MiniCPM5-1B + LoRA   (rune_goblin.inference)
+   → fine-tuned openbmb/MiniCPM-V-4.6 + LoRA   (rune_goblin.inference)
    → spell outcome JSON              (validated by rune_goblin.schema)
    → game state engine               (rune_goblin.game, clamps HP)
    → updated UI
@@ -271,7 +271,7 @@ context, then returns attack/VFX metadata for the renderer.
 
 ## Notes on the model
 
-- **Fine-tuning base**: `openbmb/MiniCPM5-1B-SFT` — safetensors, llama-arch,
+- **Fine-tuning base**: `openbmb/MiniCPM-V-4.6` — safetensors, vision-capable,
   LoRA-trainable. This is what `rune-goblin-download` and `finetune.py` use.
 - **Vision fine-tune**: `ASHu2/goblinV1` — merged MiniCPM-V model for canvas
   drawings. The local GGUF files live in `models/goblinV1-gguf/gguf/`; use
@@ -281,9 +281,11 @@ context, then returns attack/VFX metadata for the renderer.
 - **Asset planner model**: `models/MiniCPM-V-4.6-gguf/MiniCPM-V-4_6-Q8_0.gguf`
   — planned higher-quality model for attack type, palette, size, area, path,
   impact reaction, particle tags and animation timing from validated spell JSON.
-- The `MODEL` in `.env` (`openbmb/MiniCPM-o-4_5-gguf`) is a **quantized GGUF
-  multimodal** model — it's the llama.cpp/serving + optional vision path and
-  **cannot** be LoRA-fine-tuned. Download it with `--gguf` only if you need it.
+- **Dialogue / story model**: the base (non-fine-tuned) `openbmb/MiniCPM-V-4.6`
+  drives NPC dialogue and story progression — loaded as the GGUF
+  `models/MiniCPM-V-4.6-gguf/MiniCPM-V-4_6-Q4_K_M.gguf` via
+  `RG_DIALOGUE_MODEL`. It's a **quantized GGUF multimodal** model and **cannot**
+  be LoRA-fine-tuned. Download it with `--gguf` only if you need it.
 
 Renderer rule: do not generate new images during combat or exploration. Use
 model metadata to drive existing sprites, particles, overlays, CSS/canvas

progress.md

@@ -147,8 +147,6 @@ uv run --extra gguf python app/vision_app.py     # → http://localhost:7861
 Play with `RG_USE_MODEL=0` to skip the model entirely (drawings then fall back
 to the rule engine; rune-button casts are unaffected). Rune-button casts are
 always instant; the first drawing cast loads the vision model (~30s on CPU).
-`.claude/launch.json` has a `rune-goblin` preview config (model off) for quick
-UI checks.
 
 ---
 

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "rune-goblin"
 version = "1.0.0"
-description = "A fine-tuned spell-language dungeon game powered by MiniCPM5-1B."
+description = "A fine-tuned spell-language dungeon game powered by openbmb/MiniCPM-V-4.6."
 readme = "README.md"
 requires-python = ">=3.10,<3.13"
 license = { text = "Apache-2.0" }

rune_goblin_plan.md

@@ -61,10 +61,10 @@ Rune Goblin fits because the AI is the core game mechanic. The player invents sp
 
 ## 4. What We Fine-Tune
 
-We fine-tune a small OpenBMB text model, preferably:
+We fine-tune the OpenBMB vision model:
 
 ```text
-MiniCPM5-1B-SFT
+openbmb/MiniCPM-V-4.6
 ```
 
 The fine-tuned text model becomes:
@@ -86,7 +86,7 @@ The full drawing pipeline is:
 Player drawing
 → goblinV1-gguf Q4_K_M sketch reader
 → spell sketch metadata / detected rune sequence
-→ fine-tuned MiniCPM5 RuneLang spell engine
+→ fine-tuned openbmb/MiniCPM-V-4.6 RuneLang spell engine
 → validated spell outcome JSON
 → MiniCPM-V-4.6 Q8 asset planner
 → animation / attack / asset metadata JSON
@@ -347,7 +347,7 @@ Split:
 6. Use a larger model or template system to create funny spell names and flavor text.
 7. Validate JSON.
 8. Save as JSONL.
-9. Fine-tune MiniCPM5-1B-SFT with LoRA.
+9. Fine-tune openbmb/MiniCPM-V-4.6 with LoRA.
 ```
 
 ---
@@ -394,7 +394,7 @@ Gradio UI
   |
 Rune Serializer
   |
-Fine-tuned MiniCPM5-1B-SFT LoRA
+Fine-tuned openbmb/MiniCPM-V-4.6 LoRA
   |
 JSON Spell Result
   |
@@ -583,7 +583,7 @@ Modal credits can be used for:
 Local / Modal:
 1. Generate rune_spell_dataset.jsonl
 2. Upload dataset to Hugging Face
-3. Fine-tune MiniCPM5-1B-SFT using LoRA/QLoRA
+3. Fine-tune openbmb/MiniCPM-V-4.6 using LoRA/QLoRA
 4. Save LoRA adapter
 5. Push adapter to Hugging Face
 6. Load adapter in Gradio or Modal endpoint
@@ -594,13 +594,13 @@ Local / Modal:
 ### 13.3 Suggested Training Setup
 
 ```text
-Base model: OpenBMB MiniCPM5-1B-SFT
+Base model: openbmb/MiniCPM-V-4.6
 Method: LoRA / QLoRA
 Dataset size: 5k examples
 Epochs: 2–4
 Max sequence length: 1024 or 2048
 Batch size: based on GPU
-Output: rune-goblin-minicpm5-1b-lora
+Output: rune-goblin-minicpm-v-4.6-lora
 ```
 
 ---
@@ -678,7 +678,7 @@ Good output:
 
 | Badge | How We Earn It |
 |---|---|
-| Well-Tuned | Fine-tuned MiniCPM5-1B-SFT on RuneLang |
+| Well-Tuned | Fine-tuned openbmb/MiniCPM-V-4.6 on RuneLang |
 | Off-Brand | Custom Gradio UI that looks like a tiny cursed dungeon |
 | Field Notes | Blog post explaining RuneLang and training process |
 | Open Trace | Publish sample game traces / spell logs |