Fix Gradio/huggingface_hub version compatibility

- Add gradio>=4.12.0 and huggingface_hub>=0.20.0 to requirements
- Update Dockerfile to upgrade both packages before install
- Add MODEL_CARD.md and merge_adapter.py

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Dockerfile

@@ -3,6 +3,7 @@ FROM python:3.10-slim
 WORKDIR /app
 
 COPY requirements.txt .
+RUN pip install --no-cache-dir --upgrade gradio huggingface_hub
 RUN pip install --no-cache-dir -r requirements.txt
 
 COPY . .

MODEL_CARD.md

@@ -0,0 +1,88 @@
+---
+language:
+- en
+- code
+library_name: transformers
+license: apache-2.0
+tags:
+- code generation
+- python
+- qwen
+- fine-tuned
+- stack-overflow
+- coding-assistant
+---
+
+# Stack 2.9 Fine-tuned
+
+A fine-tuned version of [Qwen2.5-Coder-1.5B](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B) trained on Stack Overflow data.
+
+## Model Details
+
+- **Base Model:** Qwen/Qwen2.5-Coder-1.5B
+- **Architecture:** Transformer decoder with grouped query attention
+- **Parameters:** 1.5B
+- **Context Length:** 8,192 tokens
+- **Precision:** FP16
+- **Trained on:** Stack Overflow Q&A data
+
+## Capabilities
+
+✅ **Code Generation** — Write Python, SQL, JavaScript, and more  
+✅ **Code Completion** — Complete functions and snippets  
+✅ **Programming Help** — Debug, explain, and refactor code  
+✅ **Natural Language** — Answer questions and chat  
+
+## Usage
+
+### Python (Transformers)
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("my-ai-stack/stack-2-9-finetuned")
+tokenizer = AutoTokenizer.from_pretrained("my-ai-stack/stack-2-9-finetuned")
+
+prompt = "def quick_sort(arr):"
+inputs = tokenizer(prompt, return_tensors="pt")
+outputs = model.generate(**inputs, max_new_tokens=100)
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+### Interactive Chat
+
+```python
+# See chat.py in the repo for an interactive CLI
+```
+
+## Training Details
+
+- **Method:** LoRA fine-tuning
+- **Rank:** 8
+- **Epochs:** ~0.8
+- **Final Loss:** 0.0205
+- **Data:** Stack Overflow code Q&A
+
+## Limitations
+
+⚠️ **Training Contamination** — May occasionally repeat training examples  
+⚠️ **Small Model** — 1.5B params; larger models (7B, 32B) perform better  
+⚠️ **Single Language** — Primarily trained on Python-heavy data  
+⚠️ **No Tool Use** — This is a base model, not an agent  
+
+## Citation
+
+```bibtex
+@misc{my-ai-stack/stack-2-9-finetuned,
+  author = {Walid Sobhi},
+  title = {Stack 2.9 Fine-tuned on Stack Overflow},
+  year = {2026},
+  publisher = {HuggingFace},
+  url = {https://huggingface.co/my-ai-stack/stack-2-9-finetuned}
+}
+```
+
+## Contact
+
+- **GitHub:** [my-ai-stack/stack-2.9](https://github.com/my-ai-stack/stack-2.9)
+- **Author:** Walid Sobhi

merge_adapter.py

@@ -0,0 +1,33 @@
+from peft import PeftModel
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+import argparse
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--base-model", required=True)
+    parser.add_argument("--adapter-path", required=True)
+    parser.add_argument("--output-path", required=True)
+    args = parser.parse_args()
+    
+    print(f"Loading base model: {args.base_model}")
+    # Load without device_map to avoid the error
+    model = AutoModelForCausalLM.from_pretrained(
+        args.base_model,
+        torch_dtype=torch.float32  # Use FP32 for stability
+    )
+    tokenizer = AutoTokenizer.from_pretrained(args.base_model)
+    
+    print(f"Loading adapter: {args.adapter_path}")
+    model = PeftModel.from_pretrained(model, args.adapter_path)
+    
+    print("Merging...")
+    model = model.merge_and_unload()
+    
+    print(f"Saving to: {args.output_path}")
+    model.save_pretrained(args.output_path)
+    tokenizer.save_pretrained(args.output_path)
+    print("✅ Done!")
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -3,6 +3,10 @@
 # Core
 stack-cli>=2.9.0
 
+# Gradio UI
+gradio>=4.12.0
+huggingface_hub>=0.20.0
+
 # Training & ML
 torch>=2.0.0
 transformers>=4.35.0

src/cli/main.py

@@ -12,8 +12,8 @@ from typing import Optional
 
 # Add parent directories to path
 sys.path.insert(0, str(Path(__file__).parent))
-sys.path.insert(0, str(Path(__file__).parent.parent / "stack-2.9-eval"))
-sys.path.insert(0, str(Path(__file__).parent.parent / "stack-2.9-training"))
+sys.path.insert(0, str(Path(__file__).parent.parent / "stack" / "eval"))
+sys.path.insert(0, str(Path(__file__).parent.parent / "stack" / "training"))
 
 from model_client import create_model_client, ChatMessage
 from pattern_miner import PatternMiner

test_model.py

@@ -410,7 +410,7 @@ def run_test(model, tokenizer, test_config: Dict) -> Dict:
     start_time = time.time()
     
     # Generate completion
-    completions = generate_completion(model, tokenizer, prompt, max_tokens=max_tokens)
+    completions = generate_completion(model, tokenizer, prompt, max_new_tokens=max_tokens)
     elapsed = time.time() - start_time
     
     # Extract and check code