fchis/laravel-buildspec-training

Laravel 13.x BuildSpec → Code Training Dataset

Training data for fine-tuning code generation models to convert structured BuildSpec JSON objects into Laravel 13.x PHP files.

This dataset was used to train fchis/Laravel-13x-Qwen2.5-Coder-7B-Instruct-LoRA-Spec.

What is BuildSpec?

BuildSpec is a structured JSON format that unambiguously describes a single Laravel artifact (model, migration, controller, resource, form_request, or pest_test). Instead of asking the model to interpret natural language, you give it an exact specification:

{
  "laravel_version": "13.x",
  "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?

The spec approach shifts error type from semantic hallucinations (hard to fix) to specification gaps (compiler-catchable):

Approach	Error type	Fix method
Natural language prompt	Model invents things not asked for	Runtime debugging
BuildSpec JSON	Wrong field names, missing spec fields	Spec compiler validates before generation

Dataset Statistics

Split	Examples	Artifacts
train	49	model(14), resource(8), controller(8), form_request(10), pest_test(5), migration(4)
valid	5	mixed

Format

Each example uses the OpenAI chat format:

{
  "messages": [
    {"role": "system",    "content": "You are a Laravel 13.x PHP code generator..."},
    {"role": "user",      "content": "{...BuildSpec JSON...}"},
    {"role": "assistant", "content": "<?php\nnamespace App\\Models;\n..."}
  ]
}

Artifact Reference

model

{
  "artifact": "model",
  "class": "Post",
  "namespace": "App\\Models",
  "table": "posts",
  "has_factory": true,
  "soft_deletes": false,
  "fillable": ["title", "body", "user_id"],
  "casts": {"published_at": "datetime"},
  "relationships": [
    {"type": "BelongsTo", "model": "User", "method": "author", "foreign_key": "user_id"},
    {"type": "HasMany",   "model": "Comment", "method": "comments"},
    {"type": "BelongsToMany", "model": "Tag", "method": "tags"}
  ],
  "scopes": [
    {"name": "published", "column": "status", "value": "published"}
  ]
}

controller

{
  "artifact": "controller",
  "class": "PostController",
  "namespace": "App\\Http\\Controllers\\Api",
  "model": "Post",
  "resource": "PostResource",
  "form_request": "StorePostRequest",
  "validation_mode": "form_request",
  "eager_load": ["author", "tags"],
  "paginate": 15,
  "filters": ["status"],
  "many_to_many": {"relation": "tags", "input_key": "tag_ids"}
}

form_request (with conditional rules)

{
  "artifact": "form_request",
  "class": "StoreEventRequest",
  "namespace": "App\\Http\\Requests",
  "rules": {
    "title":      ["required", "string", "max:255"],
    "venue_id":   [],
    "event_date": ["required", "date"]
  },
  "conditional_rules": {
    "venue_id": {
      "POST": ["required", "integer", "exists:venues,id"],
      "PUT":  ["sometimes", "integer", "exists:venues,id"]
    },
    "event_date": {
      "POST": ["after:now"]
    }
  }
}

Important: Never use literal conditional tokens like "required_on_post" in rules[]. Use conditional_rules{} instead. The spec compiler rejects them.

resource

{
  "artifact": "resource",
  "class": "PostResource",
  "namespace": "App\\Http\\Resources",
  "fields": [
    {"key": "id",         "source": "id"},
    {"key": "title",      "source": "title"},
    {"key": "created_at", "source": "created_at"}
  ],
  "loaded_relations": [
    {"key": "author", "resource": "AuthorResource", "type": "make"},
    {"key": "tags",   "resource": "TagResource",    "type": "collection"}
  ]
}

pest_test

{
  "artifact": "pest_test",
  "class": "Post",
  "namespace": "Tests\\Feature",
  "model": "App\\Models\\Post",
  "endpoints": [
    {"method": "GET",    "path": "/api/posts",      "action": "index"},
    {"method": "POST",   "path": "/api/posts",      "action": "store"},
    {"method": "GET",    "path": "/api/posts/{id}", "action": "show"},
    {"method": "PUT",    "path": "/api/posts/{id}", "action": "update"},
    {"method": "DELETE", "path": "/api/posts/{id}", "action": "destroy"}
  ],
  "required_on_create": ["title", "body"]
}

Full Pipeline

See laravel-ai-gen for the complete pipeline:

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

Pipeline stages:

1. Planner (planner.py): NL description → BuildSpec JSON array (few-shot)
2. Compiler (spec_compiler.py): validates + normalizes each spec
3. Generator: each spec → PHP file (this model)
4. Syntax check: php -l on all written files

Training Details

• Base model: Qwen2.5-Coder-7B-Instruct (4-bit quantized via MLX)
• Method: LoRA (rank 8, --num-layers 8)
• Training framework: mlx-lm
• Hardware: Apple M2 Pro 16GB
• Iterations: 225 cumulative (5 rounds: 100+25+25+25+50)
• Final val_loss: 0.065
• Command:

  mlx_lm.lora \
    --model mlx-community/Qwen2.5-Coder-7B-Instruct-4bit \
    --train \
    --data data_spec \
    --adapter-path adapters_spec \
    --batch-size 1 --iters 100 \
    --learning-rate 1e-5 \
    --num-layers 8 \
    --max-seq-length 1400
  

Results

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

Metric	Score
PHP syntax valid	26/26 (100%)
Eval perfect (0 bugs)	26/26 (100%)
Pest tests pass	20/20 (100%)
Manual fixes needed	3 (SoftDeletes trait, JsonResource import, repetition loop)

Citation

@misc{laravel-buildspec-2026,
  author    = {Florinel Chis},
  title     = {Laravel 13.x BuildSpec to Code Training Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/fchis/laravel-buildspec-training}
}