SpawnPoint-Coder-7B v6: card games (deckbuilders)

README.md

@@ -20,7 +20,12 @@ What it does inside the pipeline:
 
 4-bit AWQ, runs on a single 12 GB GPU.
 
-**What's new in v5.2, v5.1 and v5**
+**What's new in v6, v5.2, v5.1 and v5**
+- **v6** designs **card games** (Slay the Spire style deckbuilders): "a deckbuilder where a wizard climbs a frozen
+  tower full of ice spirits", "game thẻ bài kiểu Slay the Spire, hiệp sĩ đánh lũ xương khô, trùm cuối là Lich". It picks
+  the hero and its class (warrior, mage, rogue), the game's own cards and what each does (big hits, sweeps, poison,
+  shields, healing...), the foes and how they fight (brutes, guards, casters, swarms) and the boss. "Upgrade cards" in
+  a survivor or "key cards" in a shooter stay what they were.
 - **v5.2** catches farm pests mentioned only in passing ("game nông trại mùa đông, có cáo", "a farm with a fox around"),
   keeps a named farmer ("a cat farmer"), and leaves a peaceful farm without pests.
 - **v5.1** fixes prompts that name the player only indirectly ("a tank shooting monsters": the tank is now the hero,
@@ -33,8 +38,8 @@ What it does inside the pipeline:
 - Fixes from v4: "a wizard shooting skeletons" keeps the shoot ability, "collect coins" is a collect goal, an unnamed
   pickup is a Coin.
 
-**Needs Spawn Point 0.3 or later**: v5 answers in the per-genre format. Older Spawn Point versions expect the v4
-format; use release `v4-release` of this repo with them.
+**Needs Spawn Point 0.4 or later for card games** (0.3 or later for the rest): v5 and later answer in the per-genre
+format. Older Spawn Point versions expect the v4 format; use release `v4-release` of this repo with them.
 
 ## Run it
 
@@ -51,37 +56,39 @@ Hand-written prompts in English and Vietnamese through the whole pipeline, same
 Sealed sets were written before training and not looked at until the release decision. The first two rows include
 prompts about the genre v4 did not know, so v4 cannot be expected to answer them.
 
-| benchmark (prompts fully right) | v4 | v5.1 | v5.2 (this release) |
+| benchmark (prompts fully right) | v4 | v5.2 | v6 (this release) |
 |---|---|---|---|
-| **sealed set: farm pests named in passing** (16 prompts) | - | 75% | 100% |
-| sealed set: shooters and the v4 fixes (16 prompts) | 56% | 75% | 88% |
-| dev set: shooters and the v4 fixes (24 prompts) | 54% | 83% | 79% |
-| sealed set: farms, tower defense, cartoon (12 prompts) | 67% | 92% | 92% |
-| 16 prompts: farms, tower defense, melee, cartoon | 75% | 94% | 81% |
-| sealed set: the first nine genres (12 prompts) | 67% | 75% | 75% |
-| 30 human-written prompts | 87% | 87% | 87% |
-| 12 prompts: bubble, rescue, survivor | 92% | 83% | 92% |
-
-**Where v5.2 is a little behind v5.1**: the 16-prompt farms / tower defense / melee set (two prompts: a "space shooter"
-read as the other kind of shooter, a barbarian's axe) and the 24-prompt shooter dev set (one prompt). Some prompts are
-genuinely ambiguous ("a space shooter" can be a side-scrolling shooter). Results vary by one to three prompts per set
-from run to run, so read differences of one or two prompts as noise.
-
-Held-out synthetic examples with one exact answer (v5 format only): briefs 77%, tweaks 98%, entity edits 100%.
+| **sealed set: card games** (16 prompts) | - | 56% | 94% |
+| sealed set: farm pests named in passing (16 prompts) | - | 100% | 100% |
+| sealed set: shooters and the v4 fixes (16 prompts) | 56% | 88% | 88% |
+| dev set: shooters and the v4 fixes (24 prompts) | 54% | 79% | 88% |
+| sealed set: farms, tower defense, cartoon (12 prompts) | 67% | 92% | 83% |
+| 16 prompts: farms, tower defense, melee, cartoon | 75% | 81% | 94% |
+| sealed set: the first nine genres (12 prompts) | 67% | 75% | 83% |
+| 30 human-written prompts | 87% | 87% | 90% |
+| 12 prompts: bubble, rescue, survivor | 92% | 92% | 75% |
+
+**Where v6 is a little behind v5.2**: the 12-prompt bubble / rescue / survivor set (two prompts, both names in a bubble
+shooter: the shooter or the egg) and the 12-prompt sealed farms / tower defense set (one prompt), and exact synthetic
+briefs (70% against 77%). Results vary by one to three prompts per set from run to run, so read differences of one or
+two prompts as noise.
+
+Held-out synthetic examples with one exact answer (v5 format only): briefs 70%, tweaks 98%, entity edits 96%.
 
 The 4-bit AWQ files you download, measured the same way:
 
 | benchmark (prompts fully right) | AWQ build |
 |---|---|
-| **sealed set: farm pests named in passing** (16 prompts) | 100% |
-| 30 human-written prompts | 83% |
-| sealed set: shooters and the v4 fixes (16 prompts) | 94% |
+| **sealed set: card games** (16 prompts) | 94% |
+| 30 human-written prompts | 87% |
+| sealed set: farm pests named in passing (16 prompts) | 100% |
+| sealed set: shooters and the v4 fixes (16 prompts) | 81% |
 
 ## Limitations
 
 - Fine-tuned only for the 2D game kinds Spawn Point builds: **platformers (including rescue missions and sword
-  fighting), top-down arenas and survivors, bubble shooters, farming games, tower defense and side-scrolling
-  shooters**. Other genres (3D, RPG systems, card games, racing…) and free-form Godot coding were not part of the
+  fighting), top-down arenas and survivors, bubble shooters, farming games, tower defense, side-scrolling
+  shooters and card games**. Other genres (3D, RPG systems, 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.
 - Prompts that do not say the genre can be read either way ("a space shooter" may come out as a side-scrolling
   shooter, "a farm game" as a farming game).

model-00001-of-00002.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ce6ca0757ea25efcfff2c3c9a46796f671b0e7242fe8275e55837ec5414c5e04
+oid sha256:1ea56ae2b3fa49289d39bd575aa5b8e341eee00681856294515d2bef0703fbef
 size 4480835056