Base model: Qwen/Qwen3-8B
Training method: GRPO with self-reflection rewards (inspired by ReflexiCoder)
Checkpoint: GRPO7 final (100 steps, group_size=12)
Learning rate: 1e-6

Reward Components

Component	Weight	Description
manifest_completeness	0.20	format_version, header, uuid, version validation
structure_building	0.20	JSON structure, module types
api_correctness	0.20	@minecraft/server API usage
js_syntax	0.20	JavaScript syntax validity
self_reflection	0.20	Correction pattern detection

Training Details

Steps: 100
Group size: 12
Reward function: Self-reflection rewards inspired by ReflexiCoder
Key insight: Lower LR (1e-6) for stability; slightly better JS API correctness (72.7% vs 72.5%)

python

from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained("alexchapin/portkit-coder-8b-grpo7")