The client needed to improve LLM performance across two core developer capabilities:

• Structured error intelligence: Existing Apex datasets lacked real-world coding mistakes and corrective logic, making it difficult for the model to identify and understand why a snippet was wrong.
• Prompt-to-query mapping depth: SOQL examples lacked linguistic diversity and contextual reasoning, limiting  the mode’s ability to translate complex user prompts into accurate and optimized SOQL or SOSL queries.

These gaps limited the model’s ability to perform context-aware reasoning, error detection, and syntactic generation, core capabilities for an enterprise-grade assistant.

Dataset

To address these challenges, the team designed a two-phase approach focused on error comprehension and semantic translation.

a. Apex Notebooks

• Built around a bad code → corrected code → explanation structure
• Sourced and validated violations using PMD (Programming Mistake Detector)
• Covered recurring anti-patterns such as hardcoded values, inefficient loops, unguarded triggers, DML misuse, and more
• Enabled fine-grained error classification, correction logic, and instruction-following training

b. SOQL Notebooks

• Each prompt simulated complex logic, including nested filters, multi-object joins, and aggregate queries
• Expert-written responses translated prompts into optimized SOQL or SOSL
• Commentary documented reasoning steps, constraints, and selection logic
• Taught the model how to reason across both syntax and business logic

Evaluation

Each notebook task underwent a structured QA process:

• Code correctness and formatting verified for Apex and SOQL responses
• Natural-language clarity and diversity reviewed for prompt quality
• Error explanations and reasoning steps checked for completeness and clarity
• Standardized notebook templates enabled scalable reuse across future training cycles

• Delivered 2,000+ Apex samples featuring real and simulated logic flaws with detailed  fix reasoning
• Delivered 2,000+ SOQL/SOSL samples aligned to complex user intent, showcasing high linguistic and structural diversity
• Achieved ~99.6% data acceptance rate for SOQL samples
• Improved LLM precision in:
  - Detecting and fixing insecure or invalid Apex logic
  - Translating prompts to accurate, optimized SOQL queries
• Provided a modular notebook format supporting future dataset expansion and fine-tuning

This dataset provided the client with a scalable, high-quality foundation for training models on real-world software development tasks. The notebooks improved:

• Error detection and fix classification
• Natural language to code translation
• Model interpretability and reasoning traceability

The client can now apply this framework to build development assistants with stronger grounding in syntax, semantics, and best practices.