SpawnPoint-Coder-7B v2: better briefs and Vietnamese, Godot model card

README.md

@@ -7,9 +7,19 @@ tags: [gamedev, godot, json, spawn-point, awq]
 
 # SpawnPoint-Coder-7B
 
-A 7B model fine-tuned for [Spawn Point](https://github.com/leluong141996-dev/Spawn-Point): it turns a game idea
-into a game brief, applies tweaks ("jump 20% higher, add a dash") and edits a single entity
-("double its health and make it shoot fireballs") as strict JSON. 4-bit AWQ, runs on a single 12 GB GPU.
+**A 7B model that generates Godot 4 games.** Fine-tuned for [Spawn Point](https://github.com/leluong141996-dev/Spawn-Point),
+an open-source autonomous game studio: you describe a game in plain English or Vietnamese, the model writes
+the game design as strict JSON, and Spawn Point compiles it into a playable Godot 4 project (scenes, GDScript
+components, levels, sprites) that runs on desktop or in the browser.
+
+What it does inside the pipeline:
+
+- **Design** — game idea → game brief (genre, character, abilities, enemies, hazards, pickups, goal, feel).
+- **Tweak** — "jump 20% higher, add a dash, more enemies" → a precise patch of the game.
+- **Edit an entity** — "double its health and make it shoot fireballs" → component and parameter changes.
+- **Translate** — Vietnamese game ideas → English briefs.
+
+4-bit AWQ, runs on a single 12 GB GPU.
 
 ## Run it
 
@@ -22,14 +32,27 @@ Then Spawn Point finds it automatically at `http://localhost:8000/v1`.
 
 ## Evaluation
 
-Exact-match rate on a held-out set of Spawn Point tasks (same prompts and JSON schemas as production):
+Held-out Spawn Point tasks (same prompts and JSON schemas as production) plus 30 human-written game ideas
+(English and Vietnamese, 153 checks):
 
-| task | base model | SpawnPoint-Coder-7B |
+| benchmark | v1 | v2 (this release) |
 |---|---|---|
-| brief | 0% | 67% |
-| tweak | 62% | 97% |
-| entity | 47% | 100% |
-| translate | 9% | 91% |
+| held-out brief (exact match) | 36% | 88% |
+| held-out tweak (exact match) | 100% | 98% |
+| held-out entity (exact match) | 100% | 91% |
+| held-out translate (exact match) | 52% | 97% |
+| human-written prompts fully right | 73% | 87% |
+| human-written prompts, checks passed | 94% | 97% |
+
+## Limitations
+
+- Fine-tuned only for **2D platformers and 2D top-down games** built with Spawn Point's component
+  library. Other genres (3D, RPG systems, strategy, card games, racing…) and free-form Godot coding were
+  not part of the training data, so expect results no better than a general 7B coder model there.
+- It writes Spawn Point's JSON formats; the GDScript itself comes from Spawn Point's tested component
+  library, not from the model. Outside Spawn Point it is a general coding model with a narrow specialty.
+- Occasionally one detail of a request is missed (e.g. a pickup or one enemy type); Spawn Point's
+  validator and normalizer catch most of these.
 
 ## License
 

model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6b9f4f3da600eccdfd10fca96c01b0f93cd071a70246bfd20c2c983a843263e0
+oid sha256:c63c4d808b401ccb67b058f73ed55fb78dc7f983f34f5d3d4085fc3b2832c797
 size 4480835056

tokenizer_config.json

@@ -205,4 +205,4 @@
   "split_special_tokens": false,
   "tokenizer_class": "Qwen2Tokenizer",
   "unk_token": null
-}
+}