migration-copilot-phi-3-5-mini-instruct

Fine-tuned Phi-3.5-mini-instruct for enterprise SQL/HiveQL/PL-SQL/Stored Procedure → PySpark migration.

Part of the Enterprise Migration Copilot project.

Benchmark Results

Evaluated on 480 held-out scripts (120 per language), never seen during training:

Language	Pass Rate
SQL	64%
HiveQL	74%
PL/SQL	57%
Stored Procedure	32%
Overall	57%

Metrics: syntax_valid AND has_pyspark_ops AND semantic_sim (60% table name coverage).

Best fine-tuned model across all 3 models trained in this project.

Training Details

• Base model: microsoft/Phi-3.5-mini-instruct (3.8B parameters)
• Method: QLoRA / LoRA fine-tuning (r=16, alpha=32)
• Training data: 1,312 validated SQL→PySpark pairs
• Data sources: 300 hand-crafted Claude examples (99.67% validation pass rate) + 1,012 Ollama-generated pairs (79.58% pass rate)
• Languages: SQL, HiveQL, PL/SQL, Stored Procedures (T-SQL)
• Epochs: 3
• Train loss: 2.133 → (high due to chat template format mismatch, but eval loss converged well)
• Eval loss: 0.294 (best across all 3 models)
• Hardware: Google Colab T4 GPU (free tier)
• Training time: ~54 minutes

Prompt Format

This model expects the exact fine-tuning prompt format:

### Instruction:
Convert the following {SOURCE_LANGUAGE} code to PySpark.
Difficulty: {difficulty}

### Input:
{source_code}

### Response:

Example

Input (PL/SQL):

DECLARE
  CURSOR c_emp IS
    SELECT emp_id, salary FROM employees WHERE dept_id = 10;
BEGIN
  FOR rec IN c_emp LOOP
    DBMS_OUTPUT.PUT_LINE(rec.emp_id || ': ' || rec.salary);
  END LOOP;
END;

Output (PySpark):

from pyspark.sql import functions as F

df = spark.table('employees')
emp_df = df.filter(F.col('dept_id') == 10).select('emp_id', 'salary')
for row in emp_df.collect():
    print(f"{row['emp_id']}: {row['salary']}")

Intended Use

• Enterprise legacy SQL migration to Apache Spark / Databricks
• Fintech data pipeline modernization
• Batch migration tooling and code review assistance

Limitations

• Stored Procedure migration (T-SQL) has lower accuracy (32%) due to complex procedural constructs
• Expert-difficulty recursive CTEs and dynamic SQL may require manual review
• Model runs via HuggingFace Inference API — cold start latency ~15-30s on free tier

Links

• 📁 GitHub: enterprise-migration-copilot
• 📊 Dataset: praveends/enterprise-migration-dataset
• 🤗 All models: praveends