Modelpublic

EditorAI-Geode/editorai-14b-v4

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face

1likes90downloads

Model Card

EditorAI v4 — 14B GD Level Designer (EAS-native)

The flagship EditorAI model. Fine-tune of Qwen/Qwen2.5-14B-Instruct trained directly on the EAS (EditorAI Script) output format.

	v3 (7B)	v4 (14B)
Base model	Qwen2.5-7B-Instruct	Qwen2.5-14B-Instruct
Native context	32 K	32 K
Q4KM size	4.46 GB	8.4 GB
Q5KM size	—	9.8 GB
Output format	EAS + JSON fallback	EAS-native (trained on EAS verbs directly)
Training data	3,700 mixed	4,173 mixed (level-gen rows pre-converted to EAS w/ auto FLOOR / SPIKE-TRAIN / PILLAR detection)

Files

—editorai-v4-Q4_K_M.gguf (8.4 GB) — ship target, recommended for ≥12 GB GPUs
—editorai-v4-Q5_K_M.gguf (9.8 GB) — quality bump if you have 16 GB+
—Modelfile.v4 — Ollama Modelfile, 32K ctx, Qwen2.5 tool template

Quick start

bash

ollama pull entity12208/editorai:v4-14b
ollama create entity12208/editorai:v4-14b -f Modelfile.v4   # alternative
./llama-server -m editorai-v4-Q4_K_M.gguf -c 32768 --jinja  # llama.cpp

Speed (Q4KM)

—RTX 3060 12 GB / 4060 Ti 16 GB: ~30–40 t/s
—RTX 4070 / 3090: ~50–70 t/s
—RTX 4090: ~80–120 t/s
—Apple M3 Max (Metal): ~25–35 t/s

Training

—QLoRA 4-bit NF4, rank 32, alpha 64, lr 2e-4 cosine, adamw_8bit
—H100 80 GB (Lightning.ai), ~1h training (260 steps × 14.6 s/step, 2 epochs)
—Gradient checkpointing on, maxlen 1024, batch 4 × gradaccum 8 (effective 32)
—4,173 rows: 2,473 EAS-native level-gen (parsed from 150 real .gmd files with smart structural macro detection) + 1,200 multi-turn tool-use + 500 Alpaca
—System prompt at training time mirrors the mod's runtime system prompt

License

Apache-2.0, inherited from the base model.