fchis/Laravel-13x-Planner-Qwen3-1.7B-LoRA

Laravel 13.x Planner — Qwen3-1.7B LoRA

A reasoning/planning model that decomposes vague Laravel feature requests into specific, actionable coding instructions with file paths. Designed as the first stage of a two-model pipeline.

How It Works

User: "simple contact form that saves to database"
         ↓
THIS MODEL (Planner, 1.7B, ~1.5s):
  <think>Form request, controller, and model.</think>
  [
    {"file": "app/Http/Requests/ContactFormRequest.php",
     "instruction": "Create a Form Request with validation: name, email, subject, message"},
    {"file": "app/Http/Controllers/ContactController.php",
     "instruction": "Write a ContactController with store method using FormRequest"},
    {"file": "app/Models/Contact.php",
     "instruction": "Create model for Contact with fillable: name, email, subject, message"}
  ]
         ↓
CODER MODEL (3B, ~1s per file) → generates PHP for each

Limitations (Honest Assessment)

• 43 training examples — covers common patterns but misses edge cases
• Sometimes omits migrations — doesn't always include database table creation
• Sometimes omits routes — despite training examples with routes
• Can over-decompose complex requests (8+ instructions when 4 would suffice)
• Instructions can be vague — saying "fillable fields" instead of listing exact fields
• The planner is the bottleneck — when it gives vague instructions, the coder produces vague output

The key insight from testing: the coder produces perfect output when given specific instructions. Improving the planner's instruction specificity is the highest-impact improvement.

Model Details

Detail	Value
Base model	Qwen3-1.7B (4-bit via MLX)
Fine-tuning	LoRA, 16 layers, mask-prompt
Training data	43 feature→instructions decompositions
Training time	~2 minutes on Apple M2 Pro 16GB
Peak memory	3.3 GB
Best val loss	0.786 (iter 50)

Companion Models

Model	Role	Link
This model	Decompose features	You're here
Coder	Generate code	fchis/Laravel-13x-Qwen2.5-Coder-3B-Instruct-LoRA
CLI tool	End-to-end pipeline	github.com/florinel-chis/laravel-ai-code-generator

Usage

from mlx_lm import load, generate
import json, re

model, tok = load("fchis/Laravel-13x-Planner-Qwen3-1.7B-LoRA")

messages = [
    {"role": "system", "content": "You are a Laravel architect. Decompose feature requests into specific coding tasks. Think briefly, then output a JSON array of objects. Each object has 'file' (path) and 'instruction' (what to write)."},
    {"role": "user", "content": "simple contact form - name, email, subject, message - saves to database"}
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
plan = generate(model, tok, prompt=text, max_tokens=1500)

match = re.search(r'\[.*?\]', plan, re.DOTALL)
tasks = json.loads(match.group()) if match else []
for t in tasks:
    print(f"  {t['file']} ← {t['instruction'][:60]}")