---
license: apache-2.0
language:
  - en
base_model:
  - Qwen/Qwen3.5-9B
---
# 🏛️ SalesforceCoder Qwen 3.5 (9B) - Structured Repository

> [!IMPORTANT]
> **REPOSITORY RENAMED:** This repository was formerly `SalesforceCoder-Qwen3.5-9B-Q4_K_M-GGUF`. All structured paths inside `files/` remain unchanged.

> [!IMPORTANT]
> **PREFERRED MODEL:** For most users, the **Q4_K_M GGUF** (located in `files/q4/`) is the recommended version. It maintains high architectural accuracy for Apex/SOQL while fitting comfortably within **6GB of VRAM**, making it ideal for local development environments.

## 📂 Repository Structure

| Path | Description |
| :--- | :--- |
| **`files/q4/`** | **Primary Quant (Q4_K_M).** Optimized for VRAM < 6GB. |
| **`files/q5/` & `files/q8/`** | High-fidelity quants for 12GB+ VRAM environments. |
| **`files/mmproj/`** | Multimodal projector GGUF for vision-capable inference. |
| **`files/f16/`** | Full model weights (Safetensors) and BF16 adapters. |
| **`files/supporting/`** | Tokenizer, chat templates, and base JSON configs. |

## 🦙 Ollama / Local Inference Quick Run

To use the preferred model with Ollama:

1. **Download the Q4 GGUF:**
   
You can manually download it from `files/q4/` or use this `curl` command:   
```
powershell
curl -L "[https://huggingface.co/aneeq-hashmi/SalesforceCoder-Qwen3.5-9B/resolve/main/files/q4/SalesforceCoder-Qwen3.5-9B.Q4_K_M.gguf](https://huggingface.co/aneeq-hashmi/SalesforceCoder-Qwen3.5-9B/resolve/main/files/q4/SalesforceCoder-Qwen3.5-9B.Q4_K_M.gguf)" -o SalesforceCoder-Qwen3.5-9B.Q4_K_M.gguf
```
2. **Prepare the Modelfile:**
   
Ensure your Modelfile is in the same directory and the FROM line points to the file you just downloaded:

```
Dockerfile
FROM ./SalesforceCoder-Qwen3.5-9B.Q4_K_M.gguf
PARAMETER num_ctx 204800
```
3. **Create & Run:**

```
PowerShell
ollama create SalesforceCoder -f Modelfile
ollama run SalesforceCoder
```
---
## 📖 Table of Contents
- [🏛️ Model Overview](#-salesforcecoder-qwen-35-9b---structured-repository)
- [📂 Repository Structure](#-repository-structure)
- [🦙 Ollama Quick Start](#-ollama--local-inference-quick-run)
- [📝 Model Information](#model-information)
  - [Technical Profile](#technical-profile)
- [🛠️ System Prompt (Apex Rules)](#system-prompt-apex-rules)
  - [Code Quality & Security](#2-code-quality--security)
  - [Testing Requirements](#3-testing-requirements)
- [⚙️ Inference & Runtime Config](#inference--runtime-config)
- [🚀 Deployment Guides](#deployment-guides)
  - [llama.cpp](#-llamacpp)
  - [LM Studio](#-lm-studio)
  - [Ollama](#-ollama)
- [🐋 Execution via Docker](#-execution-via-docker-the-no-install-way)
  - [🏛️ Architecture Note](#️-architecture-note)
- [⚖️ Acknowledgments & Licensing](#️-acknowledgments--licensing)
  - [Dataset Attribution](#dataset-attribution)
---
# Model Information

## Description
**SalesforceCoder-Qwen3.5-9B** is a fine-tuned variant of Qwen3.5-9B, purpose-built for Salesforce Enterprise Architecture, Apex development, and troubleshooting. 

The model was trained on curated Salesforce Q&A, StackExchange, and GitHub data, specifically optimized for domain-specific practical problem coverage.

It also incorporates [Gianloko's Apex Code dataset](https://huggingface.co/datasets/Gianloko/apex-coder-training-data), providing exposure to real-world Salesforce coding patterns.

### Core Strengths
- **Compilable Apex:** Generates triggers, handlers, and test classes following bulkification and governor-aware best practices.
- **Advanced Architecture:** Optimized for multi-org strategy, Service Layer patterns, and secure integration designs.
- **Deep Debugging:** Diagnoses recursion, SOQL-in-loops, and flaky tests with concrete, testable fixes.
- **Unit Testing:** Prioritizes `seeAllData=false`, `Test.startTest()/Test.stopTest()`, and robust test data builders (aiming for 90%+ coverage).
- **Security First:** Enforces CRUD/FLS checks, Named Credentials, and sanitizes dynamic SOQL.
- **200k Context:** Large window allows for ingesting entire multi-file repositories or massive debug logs for cross-file analysis.

### Technical Profile
- **Memory Footprint:** ~6 GB RAM (Q4 version).
- **Context Window:** Up to 200,000 tokens.
- **Developer:** Aneeq Hashmi
- **License:** Apache-2.0
---

# System Prompt (Apex Rules)

You are a **Salesforce Enterprise Architect**. Follow these strict rules on every response:

### 1. Role & Tone
- Act as a senior reviewer. Be concise, pragmatic, and solution-focused.
- Prioritize **"Truth over Reassurance"** regarding governor limits.

### 2. Code Quality & Security
- **Visibility:** Default to `with sharing` for all classes.
- **Security:** Enforce CRUD/FLS checks on all DML/SOQL operations.
- **Best Practices:** Ensure bulkification, SOQL outside loops, and proper exception handling.
- **Sanitization:** Use bind variables and `escapeSingleQuotes` for dynamic SOQL.

### 3. Testing Requirements
- Provide comprehensive test classes with **≥ 90% coverage**.
- Verify positive, negative, and bulk scenarios.
- Use `Test.startTest()` and `Test.stopTest()` for all asynchronous logic.

---

# Inference & Runtime Config

To ensure deterministic and syntactically correct Apex, use these parameters:

| Parameter | Recommended Value | Purpose |
| :--- | :--- | :--- |
| **Temperature** | `0.0 – 0.2` | Precise, deterministic code output. |
| **Top_P** | `0.9` | Balance between variety and relevance. |
| **Repeat Penalty** | `1.1 – 1.2` | Reduce boilerplate in long classes. |
| **Context Window** | Up to `200,000` | Support for full-org analysis. |

---

# Deployment Guides

### 🦙 llama.cpp
```bash
./main -m files/q4/SalesforceCoder-Qwen3.5-9B.Q4_K_M.gguf -c 200000 --temp 0.0 --top_p 0.9 --repeat_penalty 1.1
```
### 💻 LM Studio
1. **Import:** Move the model folder into your LM Studio models directory.
2. **Context:** Under **Hardware Settings**, set the context limit to the maximum supported by your VRAM (up to 200k).
3. **Parameters:** Set **Temperature** to `0.1` and **Repeat Penalty** to `1.1`.
4. **System Prompt:** Paste the **Enterprise Architect** system prompt from the section above into the System Instruction box.

### 🚀 Ollama
Ensure you are in the root directory where the `Modelfile` is located. This command will build the model and reference the structured paths automatically:

```bash
ollama create SalesforceCoder -f Modelfile
```
# 🐋 Execution via Docker (The "No-Install" Way)

Since you are running without a local Ollama installation, use this specific two-step command to build your model inside a container. Run this from your `C:\unsloth\source-repo` directory:

### 1. Start the Container
This maps your local folder (`${PWD}`) to the container's `/root/repo` path so it can access the weights.

```powershell
docker run -d -v ${PWD}:/root/repo -p 11434:11434 --name ollama-sf ollama/ollama
```
### 2. Create the Model
This triggers the build process using the Modelfile and the 200k context configuration.

```powerShell
docker exec -it ollama-sf ollama create SalesforceCoder -f /root/repo/Modelfile
```
### 🏛️ Architecture Note
The volume mapping (-v ${PWD}:/root/repo) is critical. It allows the container to resolve the FROM ./files/q4/... path defined in your Modelfile. Without this, the model creation will fail with a "file not found" error.

# ⚖️ Acknowledgments & Licensing

### Base Model
This model is built upon **Qwen 3.5 (9B)** by the Qwen Team, licensed under **Apache 2.0**.

### Dataset Attribution
A significant portion of the fine-tuning for this model utilized the **[Apex Coder Training Data](https://huggingface.co/datasets/Gianloko/apex-coder-training-data)** created by **Gianloko**. 

* **License:** Apache 2.0
* **Usage:** This dataset provided the foundational patterns for Apex trigger logic, bulkification, and Salesforce-specific unit testing. We are grateful to Gianloko for providing this high-quality open-source resource for the Salesforce developer community.

### Repository License
The modifications, fine-tuning configurations, and repository structure provided here are licensed under the **Apache License 2.0**. 

---
*Note: This model is an independent research project and is not affiliated with, sponsored by, or endorsed by Salesforce, Inc.*