Laravel-13x-Qwen2.5-Coder-7B-Instruct-LoRA-Spec

Laravel 13.x BuildSpec → Code (LoRA Adapter)

LoRA adapter for Qwen2.5-Coder-7B-Instruct that converts structured BuildSpec JSON objects into complete Laravel 13.x PHP files.

Base model: Qwen/Qwen2.5-Coder-7B-Instruct Quantized base for MLX: mlx-community/Qwen2.5-Coder-7B-Instruct-4bit Training dataset: fchis/laravel-buildspec-training

What is BuildSpec?

Instead of natural language, you describe each Laravel artifact precisely:

{
  "artifact": "model",
  "class": "Subscriber",
  "namespace": "App\\Models",
  "table": "subscribers",
  "has_factory": true,
  "soft_deletes": false,
  "fillable": ["email", "name", "status", "subscribed_at"],
  "casts": {"subscribed_at": "datetime"},
  "relationships": [],
  "scopes": [{"name": "active", "column": "status", "value": "active"}]
}

The model outputs the complete PHP file — no markdown, no explanations.

Why BuildSpec?

Spec errors are compiler-catchable before generation runs.

Usage (MLX)

from mlx_lm import load, generate

model, tokenizer = load(
    "mlx-community/Qwen2.5-Coder-7B-Instruct-4bit",
    adapter_path="fchis/Laravel-13x-Qwen2.5-Coder-7B-Instruct-LoRA-Spec"
)

SYSTEM = """You are a Laravel 13.x PHP code generator. Input is a BuildSpec JSON object. Output is the complete PHP file.

Rules:
- Output ONLY the PHP file. No markdown fences. No explanation.
- artifact=model: ALWAYS add use HasFactory when has_factory=true. ONLY add relationship methods listed in relationships[].
- artifact=controller: ALWAYS import App\\Http\\Controllers\\Controller and Illuminate\\Http\\Request. destroy() returns response()->noContent(). store() returns response()->json($resource, 201).
- artifact=resource: ALWAYS import Illuminate\\Http\\Resources\\Json\\JsonResource. For EVERY entry in loaded_relations[] that has a "resource" key, add the use import.
- artifact=form_request: If conditional_rules present, expand using $this->isMethod('POST') ternary or spread. For POST-only rules use spread: `...$this->isMethod('POST') ? ['rule'] : []`. authorize() returns true."""

spec = {
    "artifact": "model",
    "class": "Post",
    "namespace": "App\\Models",
    "table": "posts",
    "has_factory": True,
    "fillable": ["title", "body", "user_id"],
    "relationships": [
        {"type": "BelongsTo", "model": "User", "method": "author", "foreign_key": "user_id"}
    ]
}

import json
msgs = [
    {"role": "system", "content": SYSTEM},
    {"role": "user", "content": json.dumps(spec, indent=2)},
]
prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
prompt += "<?php\n"
result = generate(model, tokenizer, prompt=prompt, max_tokens=1000, verbose=False)
print("<?php\n" + result)

Supported Artifacts

Full Pipeline

See laravel-ai-gen:

# NL → specs → compile → generate PHP files
python3 pipeline_spec.py "Create a REST API for managing blog posts with tags"

Training Details

• Base model: Qwen2.5-Coder-7B-Instruct (4-bit quantized via MLX)
• Method: LoRA (rank 8, --num-layers 8)
• Framework: mlx-lm
• Hardware: Apple M2 Pro 16GB
• Training data: 49 examples (45 train / 4 valid)
• Iterations: 175 total (100 base + 25 v2 + 25 v3 + 25 v4)
• Final val_loss: 0.031

Results

Tested on a 3-app benchmark (Subscriber API, Book Library, Event Management):

Comparison: Spec vs Prompt approach

Spec fixes are deterministic and compiler-catchable. Prompt fixes require runtime debugging.

Citation

@misc{laravel-buildspec-2026,
  author    = {Florinel Chis},
  title     = {Laravel 13.x BuildSpec to Code LoRA Adapter},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/fchis/Laravel-13x-Qwen2.5-Coder-7B-Instruct-LoRA-Spec}
}