A LoRA adapter that teaches Qwen3-Coder-30B-A3B-Instruct to write Verse, the
programming language of Unreal Editor for Fortnite (UEFN). Verse is scarce in
open pre-training data, so base coder models default to Python/C#-shaped guesses for
Verse prompts. This adapter fixes that — and it runs on a single RTX 4090.
Headline: on a compile-gated benchmark (real UEFN compiler, raw first-pass, no
retries), the adapter passes 20% vs the identical base's 4% — a +400%
relative improvement at the same quantization and hardware.
What it's for
Generating compilable Verse for UEFN gameplay: device scripts (creative_device),
scene-graph components, and HUD/widget code from a plain-English task + the devices
involved. It was built to power a Verse learning/authoring pipeline, and this V1 is
released so the community can run a capable Verse model locally.
Data: ~900 supervised pairs + a raw Verse corpus — all compile-verified or
first-party: device/API reference articles whose examples passed the real UEFN
compiler, an API-surface Q/A set built from the UEFN digests, and Verse source.
No scraped/unverified code.
Final train loss ≈ 0.95 (from ~2.1), token-accuracy ≈ 0.79.
Evaluation (the honest version)
Every candidate script is compiled on the real UEFN compiler. Metric = raw
first-pass compile-pass rate, no escalation, no retries — the hardest, least-flattering
bar. Test set = 50 device-diverse craft tasks spanning three paradigms (device-verse /
scene-graph / widget); the numbers below are n=25.
The key control — same Q4 endpoint, LoRA on vs off (scale 0):
compile-pass (raw first-pass, n=25)
Verse-Coder-30B-v1 (LoRA ON)
20% (5/25)
Qwen3-Coder-30B base (LoRA scale 0, same endpoint)
4% (1/25)
Δ (the adapter's contribution)
+16 pts / +400% relative
By paradigm (LoRA on): device-verse 22% · scene-graph 25% · widget 12% — it generalizes
past devices, not a device-only model.
For scale: on this same harness, Claude Sonnet passes ~80–100% (the frontier ceiling),
and a much larger production 35B base scores ~20% — i.e. this 4090-sized adapter matches a
model class above its weight on Verse specifically.
What the adapter actually learned (one example)
A representative base failure — it doesn't know Verse's class syntax and writes it C#/Java-style:
# BASE (LoRA off):
class CountdownGame extends creative_device # ← wrong language shape
→ Script error 3100: Unexpected "CountdownGame"
Caveat, stated plainly: this is a Q4_K_M quant on a single 4090 — the
accessible-hardware configuration, not a quality ceiling. fp16/higher-quant serving
is expected to score higher. 20% raw-first-pass is a floor; with a single
compiler-error fix loop, ~half the near-misses (err=1) resolve.
Usage
Merge or apply the adapter to the base, or run the included GGUF (~52 MB) with
llama.cpp against a Qwen3-Coder-30B-A3B GGUF:
Prompt with a clear task + the UEFN devices involved. Ground it in real device APIs where
you can — the model is strongest when told the exact device methods/events to use.
Limitations
Verse and UEFN evolve; APIs drift. Always compile in UEFN.
Q4 first-pass ~20% — treat output as a strong draft to compile-check + fix, not
guaranteed-correct code.
Trained on gameplay-device Verse; niche APIs (advanced UI, scene-graph edge cases) are weaker.
License & attribution
Adapter released under Apache-2.0, matching the base
Qwen/Qwen3-Coder-30B-A3B-Instruct. "Verse", "UEFN", and "Fortnite" are trademarks of
Epic Games; this is an independent community model, not affiliated with or endorsed by
Epic Games. Built by Verse Island — a Biloxi Studios Inc project (biloxistudios.com).
