migration-copilot-deepseek-coder-1-3b-instruct

Fine-tuned deepseek-coder-1.3b-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	29%
HiveQL	32%
PL/SQL	27%
Stored Procedure	20%
Overall	27%

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

Training Details

• Base model: deepseek-ai/deepseek-coder-1.3b-instruct (1.3B parameters)
• Method: 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: 0.347 → 0.258 (best training convergence across all 3 models)
• Eval loss: 0.363 → 0.329
• Hardware: Google Colab T4 GPU (free tier)
• Training time: ~19 minutes (fastest training due to smallest model size)

Prompt Format

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

### Input:
{source_code}

### Response:

Example

Input (HiveQL):

SELECT user_id, tag
FROM user_tags
LATERAL VIEW EXPLODE(tags) t AS tag
WHERE size(tags) > 0
DISTRIBUTE BY user_id
SORT BY tag;

Output (PySpark):

from pyspark.sql import functions as F

result = (
    user_tags
    .filter(F.size('tags') > 0)
    .select('user_id', F.explode('tags').alias('tag'))
    .repartition('user_id')
    .sortWithinPartitions('tag')
)
result.show()

Intended Use

• Enterprise legacy SQL migration to Apache Spark / Databricks
• Lightweight/edge deployment (1.3B params, smallest memory footprint of the 3 models)
• Fast prototyping and experimentation

Limitations

• Lowest overall accuracy (27%) due to smallest base model size
• Best training loss convergence (0.258) but smallest base model capacity limits output quality
• Stored Procedure migration accuracy is 20% — complex T-SQL constructs require manual review

Model Comparison

Model	Params	Overall	Train Loss	Eval Loss
deepseek-coder-1.3b (this)	1.3B	27%	0.258	0.329
qwen2.5-coder-1.5b	1.5B	45%	0.307	0.344
phi-3.5-mini	3.8B	57%	2.133	0.294

Links

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