Qwen2.5-Coder-1.5B-Odoo-Adapter.gguf

license: mit base_model: Qwen/Qwen2.5-Coder-1.5B tags:

• text-generation-inference
• lora
• gguf
• odoo
• qwen
• code
• code-generation

---

Qwen2.5-Coder-1.5B-Odoo-Adapter-GGUF

A high-precision, low-rank adaptation (LoRA) adapter specifically fine-tuned for the Odoo 19.0 enterprise ecosystem. This model acts as an autocomplete and coding companion engine designed to seamlessly understand Odoo-specific design patterns, inheritance frameworks, dynamic views layouts, multi-relational field mechanisms, and API decorators.

This adapter is distributed in the GGUF format for native consumption with llama.cpp utilities, enabling ultra-fast, local, and completely offline code-completions right inside your favorite IDE.

---

📊 Model & Training Features

The adapter was baked using packed Sequence Fine-Tuning (SFT) over a multi-archive codebase compilation, exposing the network to structural programming sequences rather than generic syntax boilerplate.

• Base Architecture: Qwen/Qwen2.5-Coder-1.5B
• Dataset Scope: Dense data extraction across Odoo 19.0 core logic, partner-contact, server-tools, and server-ux repositories (~127MB raw codebase footprint, aggressively scrubbed of redundant token bleed).
• Target Configurations: * Full Comprehensive Attention + MLP Layer Interception (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj)
  • Rank ($r$): 16
  • Alpha ($\alpha$): 32
  • Context Window Configuration ($c$): Expanded context limits (Optimized for up to 16,384 tokens)
• Training Convergence Metrics (Epoch 3 Peak):
  • Final Training Loss: 0.5939
  • Next-Token Prediction Accuracy: 85.42%
  • Gradient Stability: Maintained standard grad_norm configurations between 0.07 and 0.11 without exploding variances.

---

🛠️ Local Implementation Workflow

Because this is a standalone GGUF LoRA adapter, it must be loaded at runtime dynamically on top of a quantized baseline Qwen model binary.

1. Prerequisites

Ensure you have an existing quantized compilation of the base model on your machine. We recommend the 8-bit quantized baseline:

• File name example: Qwen2.5-Coder-1.5B.Q8_0.gguf

2. Runtime Server Deployment

You can use llama-server to run a persistent backend endpoint that handles token predictions. Execute the command inside your local execution tree:

./llama-server.exe `
  -m "C:\ai\models\Qwen2.5-Coder-1.5B.Q8_0.gguf" `
  --lora "C:\ai\models\Qwen2.5-Coder-Odoo-Adapter.gguf" `
  -c 5120 `
  -ngl -1 `
  -fa on `
  --port 8080