Spaces:
Sleeping
Sleeping
Commit History
-
Harden vision cast recovery
Unverified signatureCo-authored-by: Codex <codex@openai.com>8d36aa9 -
Add minimap and biome visual polish to the RPG
Unverified signatureCo-authored-by: Codex <codex@openai.com>5a28fee
-
Remove legacy vision training scripts
Unverified signatureCo-authored-by: Codex <codex@openai.com>82c2019 -
Add dataset exploration notebook
Unverified signatureCo-authored-by: Codex <codex@openai.com>975e928 -
Add vision fine-tuning doc + dataset prep for MiniCPM-V 4.6
Unverified signature- docs/FINETUNE_VISION.md: end-to-end guide to fine-tune the visual spell reader — GGUF-vs-safetensors caveat, dataset analysis (5k 256x256 canvases, visual_reading+spell JSON schema), ms-swift + LLaMA-Factory LoRA recipes, 16GB VRAM guidance, eval metrics, and GGUF export for game serving. - scripts/prepare_vision_dataset.py: converts the provided splits to absolute image paths in ms-swift and LLaMA-Factory ShareGPT formats (verified 4500/500).cccc5fb -
Scaffold Rune Goblin: RuneLang engine, fine-tuning, game UIs
Unverified signature- src/rune_goblin: RuneLang vocab/grammar/combos/enemies, Pydantic spell schema + JSON repair, deterministic spell-physics engine (oracle + runtime fallback + HP clamping), synthetic dataset generator, LoRA/QLoRA finetune (TRL+PEFT), model-or-rules inference, 5-room game state machine, game-engine eval metrics. - app/app.py: Gradio dungeon UI (HF Space deliverable). - api/ + frontend/: FastAPI backend + Vite/React custom cursed-dungeon UI. - pyproject.toml (uv) with [train] extra; configs/, scripts/modal_train.py. - data/rune_goblin_dataset/: prepared 5k dataset (4500/500, 100% schema-valid). Verified: base model downloaded, full inference path, and a 5-step QLoRA training run on the dataset all work on the local GPU.7b007ed