Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,61 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen2.5-Coder-14B-Instruct
+tags:
+  - unreal-engine
+  - ue5
+  - game-development
+  - gguf
+  - qwen2
+language:
+  - en
+pipeline_tag: text-generation
+---
+
+# ue-expert-v2 (Qwen2.5-Coder-14B-Instruct + SFT)
+
+Fine-tuned Unreal Engine 5 expert model, specialized in C++ and Blueprint development.
+
+## Model Details
+
+- **Base model:** [Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct)
+- **Fine-tuning:** QLoRA (rank 32, alpha 64) SFT on 27,738 curated UE5 Q&A pairs
+- **Negative examples:** 9.3% of training data teaches the model to say "I don't know" for hallucinated/non-existent APIs
+- **Quantization:** Q4_K_M (4.87 bits per weight)
+- **Size:** ~8.4 GB GGUF
+
+## Training Data
+
+- 27,738 training pairs covering UE5 C++ APIs, Blueprint patterns, architecture, and best practices
+- 7 template categories for positive examples (hierarchy, API lookup, code patterns, etc.)
+- Negative examples include fabricated class names, non-existent functions, and plausible-but-wrong API claims
+- Quality-gated: all pairs scored >= 0.4 by automated quality pipeline
+
+## Usage with Ollama
+
+```bash
+# Download the GGUF and Modelfile, then:
+ollama create ue-expert -f Modelfile
+ollama run ue-expert "What is the parent class of ACharacter?"
+```
+
+## Files
+
+- `model-q4_k_m.gguf` — Quantized model (Q4_K_M, 8.4 GB)
+- `sft_adapter/` — LoRA adapter weights (for further fine-tuning)
+- `Modelfile` — Ollama model definition with ChatML template
+
+## Training Metrics
+
+| Metric | Value |
+|--------|-------|
+| Steps | 1299 (3 epochs) |
+| Final train loss | 0.668 |
+| Final eval loss | 0.677 |
+| Hardware | RunPod A100 SXM 80GB |
+| Training time | ~3h 46m |
+| VRAM usage | 15.8 GB / 80 GB |
+
+## Part of game-dev-docs
+
+This model is the synthesis layer for a RAG + fine-tuned model + MCP server pipeline that provides deep Unreal Engine awareness to Claude Code. The RAG pipeline provides retrieval over 302K indexed documentation chunks; this model provides internalized domain knowledge for synthesis and judgment calls.