Upload README.md with huggingface_hub

README.md

@@ -0,0 +1,150 @@
+---
+language:
+- en
+license: apache-2.0
+base_model: Qwen/Qwen2.5-Coder-3B-Instruct
+tags:
+- salesforce
+- apex
+- lwc
+- lightning-web-components
+- code
+- fine-tuned
+- gguf
+- unsloth
+pipeline_tag: text-generation
+---
+
+# DeepForce Coder v2
+
+A Salesforce-specialized AI coding assistant fine-tuned on Qwen 2.5 Coder 3B.
+Built specifically for Salesforce developers to generate, debug, review, and refactor
+Apex code and Lightning Web Components following enterprise best practices.
+
+> ✅ **v2 improvements over v1:**
+> - Fixed hallucinated Apex APIs
+> - Simple requests now generate clean minimal code
+> - Improved recursion guard patterns
+> - Custom weighted adapter merge for better task balance
+> - specialized adapters trained independently
+
+## Capabilities
+
+| Task | Description |
+|------|-------------|
+| Apex Generation | Write production-ready Apex classes, triggers, batch, scheduled, queueable |
+| LWC Development | Create Lightning Web Components with SLDS conventions |
+| Code Debug | Identify bugs with severity ratings and corrections |
+| Code Review | Review code against Salesforce best practices |
+| Refactoring | Simplify over-engineered code while preserving security |
+| Test Classes | Generate comprehensive Apex test classes |
+| Recursion Guards | Correct Apex recursion guard patterns |
+| Simple Apex | Clean minimal Apex for simple requirements |
+
+## Best Practices Enforced
+
+- `with sharing` on all classes
+- `WITH USER_MODE` on all SOQL queries
+- `Security.stripInaccessible()` before DML
+- `try-catch` on all DML and callouts
+- `Database.update/insert(records, false)` for bulk DML
+- No SOQL or DML inside loops
+- Bulkified trigger handlers with recursion guards
+- Static Set<Id> recursion guard pattern
+
+## Model Details
+
+- **Base model**: Qwen/Qwen2.5-Coder-3B-Instruct
+- **Fine-tuning**: specialized LoRA adapters with custom weights
+- **Training data**: curated Salesforce-specific examples generated via Claude API
+- **Quantization**: Q4_K_M GGUF (1.80 GB)
+- **Context length**: 6144 tokens
+
+## Quick Start
+
+### Ollama
+```bash
+ollama run hf.co/deepforce/deepforce-coder-v2:Q4_K_M
+```
+
+### llama.cpp
+```bash
+llama serve -hf deepforce/deepforce-coder-v2:Q4_K_M
+```
+
+### Python (llama-cpp-python)
+```python
+from llama_cpp import Llama
+
+llm = Llama.from_pretrained(
+    repo_id  = "deepforce/deepforce-coder-v2",
+    filename = "deepforce-coder-v2-q4_k_m.gguf",
+)
+
+response = llm.create_chat_completion(messages=[
+    {"role": "system", "content": "You are DeepForce Coder, an expert Salesforce developer."},
+    {"role": "user",   "content": "Write a simple Apex class that returns Accounts by industry."}
+])
+print(response["choices"][0]["message"]["content"])
+```
+
+## Example Prompts
+
+**Generate Apex:**
+Write a trigger handler for Opportunity that creates a follow-up Task
+
+when StageName changes to Closed Won.
+
+**Debug Apex:**
+Debug the following Apex code: [paste your code]
+
+**Review Apex:**
+Review the following Apex code for best practices: [paste your code]
+
+**Generate LWC:**
+Create an LWC component that displays a list of Accounts in a lightning-datatable.
+
+**Refactor Apex:**
+Refactor the following Apex code to the minimum implementation: [paste your code]
+
+**Generate Test Class:**
+Generate a comprehensive test class for the following Apex class: [paste your code]
+
+## Adapter Architecture
+
+DeepForce Coder v2 uses a weighted combination of 9 independently trained LoRA adapters:
+
+| Adapter | Weight | Purpose |
+|---------|--------|---------|
+| apex-main | 18% | Complex Apex generation |
+| apex-simple | 18% | Simple clean Apex patterns |
+| lwc | 15% | Lightning Web Components |
+| apex-testclass | 12% | Test class generation |
+| apex-refactor | 10% | Code refactoring |
+| apex-recursion | 10% | Recursion guard patterns |
+| apex-debug | 8% | Bug identification |
+| apex-review | 5% | Code review |
+| apex-basic | 4% | Basic patterns |
+
+## Known Limitations in v2
+- Test class generation occasionally uses System.assertEquals instead of Assert class
+- Recursion guard ID extraction needs improvement (fixed in v3)
+- These will be addressed in v3
+
+## Training
+
+Fine-tuned using [Unsloth](https://github.com/unslothai/unsloth) on Google Colab L4 GPU.
+Training data generated using Anthropic Claude API.
+Each adapter trained independently with set_seed(42) for reproducibility.
+
+## Version History
+
+| Version | Status | Notes |
+|---------|--------|-------|
+| v1 | ⚠️ Superseded | Hallucinated APIs, over-engineered simple requests |
+| v2 | ✅ Current | Fixed hallucinations, custom weighted merge |
+| v3 | 🔄 Planned | Assert class fix, better recursion guards |
+
+## License
+
+Apache 2.0 — free for commercial and personal use.