Laravel-13x-Llama-3.2-1B-Instruct-LoRA

Laravel 13.x Llama-3.2-1B-Instruct LoRA

A Llama 3.2 1B model fine-tuned on Laravel 13.x instruction-to-code pairs. Generates Laravel PHP code from natural language instructions.

This is a companion model to fchis/Laravel-13x-Qwen2.5-Coder-3B-Instruct-LoRA, created as part of a research investigation comparing different base models and training data formats for framework-specific code generation.

Key Results

Compared against yannelli/Laravel-11-Llama-3.2-1B (trained on 14K doc Q&A pairs) and base Qwen2.5-Coder-14B (no fine-tuning):

Highlights

• Middleware output was PERFECT - correct nullsafe operator, 403 JSON response, no hallucination
• Migration output was PERFECT - modern anonymous class, correct soft deletes, cascading FK
• 90 code pairs beat 14,000 doc Q&A pairs - training data format matters more than quantity

Model Details

Training Data

Dataset: fchis/Laravel-13x-Code-Instructions

90 instruction-to-code pairs from Laravel 13.x official documentation covering: Eloquent models, relationships, controllers, migrations, validation, Form Requests, middleware, views, Blade templates, View Composers, API Resources, routing, and HTTP Requests.

Usage with MLX

from mlx_lm import load, generate

model, tokenizer = load("fchis/Laravel-13x-Llama-3.2-1B-Instruct-LoRA")

messages = [
    {"role": "system", "content": "You are a senior Laravel developer. Write clean, production-ready Laravel code."},
    {"role": "user", "content": "Write a Laravel middleware that checks if the user has an active subscription and returns 403 JSON if not."}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
response = generate(model, tokenizer, prompt=text, max_tokens=800)
print(response)

Example Outputs

Middleware (PERFECT)

Prompt: "Write a Laravel middleware that checks if the authenticated user has an active subscription. If not, return a 403 JSON response."

<?php

namespace App\Http\Middleware;

use Closure;
use Illuminate\Http\Request;
use Symfony\Component\HttpFoundation\Response;

class EnsureUserIsSubscribed
{
    public function handle(Request $request, Closure $next): Response
    {
        if (! $request->user()?->hasActiveSubscription()) {
            return response()->json([
                'message' => 'You do not have an active subscription.'
            ], 403);
        }

        return $next($request);
    }
}

Migration (PERFECT)

Prompt: "Write a migration for products table with slug, price, FK, soft deletes, timestamps."

<?php

use Illuminate\Database\Migrations\Migration;
use Illuminate\Database\Schema\Blueprint;
use Illuminate\Support\Facades\Schema;

return new class extends Migration
{
    public function up(): void
    {
        Schema::create('products', function (Blueprint $table) {
            $table->id();
            $table->string('name');
            $table->string('slug')->unique();
            $table->unsignedInteger('price');
            $table->text('description')->nullable();
            $table->foreignId('category_id')->constrained()->cascadeOnDelete();
            $table->softDeletes();
            $table->timestamps();
        });
    }

    public function down(): void
    {
        Schema::dropIfExists('products');
    }
};

Form Request (NEAR PERFECT)

Prompt: "Create a Form Request for storing a blog post with validation."

<?php

namespace App\Http\Requests;

use Illuminate\Foundation\Http\FormRequest;

class StoreBlogPostRequest extends FormRequest
{
    public function authorize(): bool
    {
        return true;
    }

    public function rules(): array
    {
        return [
            'title' => ['required', 'string', 'max:255'],
            'body' => ['required', 'string'],
            'category_id' => ['required', 'string', 'must exist in categories table'],
            'tags' => ['nullable', 'array'],
            'tags.*' => ['exists:tags,id'],
            'published_at' => ['nullable', 'date', 'after:now'],
        ];
    }
}

Note: One error - 'must exist in categories table' should be 'exists:categories,id'

Training Loss

Best checkpoint at iteration 200 (lowest validation loss). The model used the iteration 500 checkpoint for these results; iteration 200 would likely score slightly higher.

Comparison: This Model vs Qwen2.5-Coder-3B

Both models trained on the same 90 examples:

Research Findings

This model was part of a broader investigation. Key findings:

1. Training data format > quantity: 90 instruction-to-code pairs dramatically outperformed 14,000 doc-Q&A pairs (yannelli's approach)
2. A 1B model can produce perfect Laravel code for well-scoped tasks (single class, single concept)
3. 3.6 GB peak memory means anyone with a Mac can fine-tune this
4. 4 minutes of training on consumer hardware produces meaningful results
5. The same approach on a code-specialized base (Qwen) gives better generalization on complex tasks

Related Models

• fchis/Laravel-13x-Qwen2.5-Coder-3B-Instruct-LoRA - Same dataset, code-specialized base, better at complex tasks
• fchis/Laravel-13x-Code-Instructions - Training dataset
• yannelli/Laravel-11-Llama-3.2-1B-Instruct - Same base model, doc-Q&A training (our benchmark comparison)

Limitations

• Only 90 training examples
• Complex multi-concept tasks still challenging
• Sometimes outputs string literals instead of validation rules (e.g., 'must exist in categories table')
• Laravel 13.x documentation only
• MLX format - requires Apple Silicon Mac
• Not a code-specialized base model - Qwen2.5-Coder variant recommended for production use