| --- |
| |
| base_model: codellama/CodeLlama-7b-Instruct-hf |
| library_name: peft |
| pipeline_tag: text-generation |
| tags: |
|
|
| - base_model:adapter:codellama/CodeLlama-7b-Instruct-hf |
| - lora |
| - transformers |
| - sql |
| - text-to-sql |
| - code-generation |
|
|
| --- |
| |
| Model Card for my-sql-model |
|
|
| Model Details |
|
|
| Model Description |
|
|
| "my-sql-model" is a SQL query generation model based on "CodeLlama-7b-Instruct-hf". |
|
|
| The model uses LoRA (Low-Rank Adaptation) through the PEFT (Parameter-Efficient Fine-Tuning) framework to adapt the base CodeLlama model for SQL generation tasks. |
|
|
| The model is intended to convert natural-language questions and database schema information into SQL queries. |
|
|
| - Developed by: tamilanda |
| - Model type: CodeLlama 7B Instruct with a LoRA adapter |
| - Language(s): English |
| - License: Refer to the base model license and repository license |
| - Base model: "codellama/CodeLlama-7b-Instruct-hf" |
| - Fine-tuning method: LoRA / PEFT |
| - Framework: Hugging Face Transformers and PEFT |
|
|
| Model Sources |
|
|
| - Repository: "tamilanda/my-sql-model" |
| - Base Model: "codellama/CodeLlama-7b-Instruct-hf" |
|
|
| Uses |
|
|
| Direct Use |
|
|
| The model can be used for natural-language-to-SQL query generation. |
|
|
| A typical input consists of: |
|
|
| 1. Database type |
| 2. Database schema |
| 3. Natural-language question |
|
|
| The model generates an SQL query corresponding to the requested operation. |
|
|
| Example: |
|
|
| Database: MySQL |
|
|
| Schema: |
| employees(id, name, department, salary) |
|
|
| Question: |
| Find employees whose salary is greater than 50000. |
|
|
| Expected output: |
|
|
| SELECT * |
| FROM employees |
| WHERE salary > 50000; |
|
|
| Downstream Use |
|
|
| The model can be integrated into: |
|
|
| - Natural-language database assistants |
| - Text-to-SQL applications |
| - RAG-based database systems |
| - Database analytics assistants |
| - Conversational SQL systems |
| - Automated SQL query generation pipelines |
|
|
| A production architecture can combine the model with schema retrieval and query validation: |
|
|
| User Question |
| ↓ |
| Database Detection |
| ↓ |
| Schema Retrieval |
| ↓ |
| Relevant Tables |
| ↓ |
| my-sql-model |
| ↓ |
| SQL Generation |
| ↓ |
| SQL Validation |
| ↓ |
| Database Execution |
| |
| Out-of-Scope Use |
|
|
| The model should not be used as an unrestricted database execution system. |
|
|
| Generated SQL should not be executed directly against production databases without validation and appropriate permissions. |
|
|
| The model is not intended for: |
|
|
| - Unauthorized database access |
| - Bypassing database permissions |
| - Destructive database operations without validation |
| - Automatic execution of untrusted SQL |
| - Security-sensitive database administration without human oversight |
|
|
| Bias, Risks, and Limitations |
|
|
| The model is a generative language model and may generate incorrect or syntactically invalid SQL. |
|
|
| Potential limitations include: |
|
|
| - Incorrect table or column selection |
| - Incorrect joins |
| - Incorrect filtering conditions |
| - Hallucinated columns or tables |
| - SQL dialect incompatibility |
| - Incorrect interpretation of ambiguous questions |
| - Incorrect aggregation or grouping |
| - Poor performance when the provided schema is incomplete |
|
|
| Generated queries should therefore be validated before execution. |
|
|
| Recommendations |
|
|
| For production applications: |
|
|
| - Provide the relevant database schema to the model. |
| - Clearly specify the database dialect. |
| - Validate generated SQL before execution. |
| - Use read-only database credentials where possible. |
| - Restrict database permissions. |
| - Apply query timeout and resource limits. |
| - Log generated queries and execution results. |
| - Require human approval for destructive operations. |
|
|
| How to Get Started with the Model |
|
|
| Install the required libraries: |
|
|
| pip install transformers peft torch |
|
|
| Load the base model and LoRA adapter: |
|
|
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| from peft import PeftModel |
| import torch |
|
|
| base_model = "codellama/CodeLlama-7b-Instruct-hf" |
| adapter_model = "tamilanda/my-sql-model" |
|
|
| tokenizer = AutoTokenizer.from_pretrained(base_model) |
|
|
| model = AutoModelForCausalLM.from_pretrained( |
| base_model, |
| torch_dtype=torch.float16, |
| device_map="auto" |
| ) |
| |
| model = PeftModel.from_pretrained( |
| model, |
| adapter_model |
| ) |
|
|
| prompt = """ |
| Generate a SQL query. |
|
|
| Database: MySQL |
|
|
| Schema: |
| employees(id, name, department, salary) |
|
|
| Question: |
| Find employees whose salary is greater than 50000. |
|
|
| Return only the SQL query. |
| """ |
|
|
| inputs = tokenizer(prompt, return_tensors="pt").to(model.device) |
| |
| with torch.no_grad(): |
| outputs = model.generate( |
| **inputs, |
| max_new_tokens=150, |
| temperature=0.1, |
| do_sample=False |
| ) |
| |
| result = tokenizer.decode( |
| outputs[0], |
| skip_special_tokens=True |
| ) |
| |
| print(result) |
|
|
| Training Details |
|
|
| Training Data |
|
|
| The model is intended for SQL query generation. The exact training dataset and dataset size should be documented separately if available. |
|
|
| A suitable training example contains: |
|
|
| Natural Language Question |
| + |
| Database Schema |
| + |
| Expected SQL Query |
|
|
| Example: |
|
|
| { |
| "question": "Find all customers from Chennai", |
| "schema": "customers(id, name, city)", |
| "query": "SELECT * FROM customers WHERE city = 'Chennai';" |
| } |
|
|
| Training Procedure |
|
|
| The model uses Parameter-Efficient Fine-Tuning (PEFT) with LoRA on the CodeLlama-7B-Instruct base model. |
|
|
| Preprocessing |
|
|
| Training examples should be formatted as instruction-following examples containing the database schema, user question, and target SQL query. |
|
|
| Training Hyperparameters |
|
|
| - Training regime: Not specified |
| - Fine-tuning method: LoRA |
| - Framework: PEFT |
| - PEFT version: 0.17.1 |
| - Base model: "codellama/CodeLlama-7b-Instruct-hf" |
|
|
| Speeds, Sizes, Times |
|
|
| Training hardware, training duration, throughput, checkpoint size, and compute requirements are not specified. |
|
|
| Evaluation |
|
|
| Testing Data, Factors & Metrics |
|
|
| Testing Data |
|
|
| The evaluation dataset is not specified. |
|
|
| Factors |
|
|
| Evaluation can be performed across: |
|
|
| - Simple SQL queries |
| - Filtering |
| - Aggregation |
| - GROUP BY |
| - ORDER BY |
| - JOIN operations |
| - Subqueries |
| - Nested queries |
| - Multiple-table queries |
| - Complex analytical queries |
|
|
| Metrics |
|
|
| Recommended metrics include: |
|
|
| - Exact Match Accuracy |
| - Execution Accuracy |
| - SQL Syntax Validity |
| - Query Execution Success Rate |
|
|
| No evaluation scores are claimed here because verified results were not provided. |
|
|
| Results |
|
|
| Evaluation results are not currently specified. |
|
|
| Summary |
|
|
| The model should be evaluated against a held-out SQL dataset before production deployment. |
|
|
| Model Examination |
|
|
| The model can be examined by testing generated SQL against known database schemas and comparing generated queries with reference queries and execution results. |
|
|
| Environmental Impact |
|
|
| Carbon emissions depend on the hardware and infrastructure used during fine-tuning. |
|
|
| - Hardware Type: Not specified |
| - Hours used: Not specified |
| - Cloud Provider: Not specified |
| - Compute Region: Not specified |
| - Carbon Emitted: Not specified |
|
|
| Carbon emissions can be estimated using the "Machine Learning Impact calculator" (https://mlco2.github.io/impact#compute). |
|
|
| Technical Specifications |
|
|
| Model Architecture and Objective |
|
|
| The model uses: |
|
|
| CodeLlama-7B-Instruct |
| ↓ |
| LoRA |
| ↓ |
| PEFT Adapter |
| ↓ |
| my-sql-model |
| |
| The objective is to adapt the base code-generation model for SQL query generation. |
|
|
| Compute Infrastructure |
|
|
| The exact training infrastructure is not specified. |
|
|
| Hardware |
|
|
| Not specified. |
|
|
| Software |
|
|
| The model uses the Hugging Face ecosystem, including: |
|
|
| - Transformers |
| - PEFT |
| - LoRA |
| - PyTorch |
|
|
| PEFT version: "0.17.1" |
|
|
| Citation |
|
|
| If this model is used in a project, cite the model repository and the underlying CodeLlama model. |
|
|
| Base Model |
|
|
| CodeLlama: Open Foundation Models for Code. |
| Meta AI. |
|
|
| Glossary |
|
|
| PEFT: Parameter-Efficient Fine-Tuning, a method for adapting large models while training a relatively small number of parameters. |
|
|
| LoRA: Low-Rank Adaptation, a PEFT technique that trains low-rank adapter matrices instead of updating all base-model parameters. |
|
|
| Text-to-SQL: Conversion of a natural-language question into an SQL query. |
|
|
| Schema: The structure of a database, including tables, columns, relationships, and data types. |
|
|
| Execution Accuracy: Measures whether the generated SQL produces the correct result when executed against the target database. |
|
|
| More Information |
|
|
| The model can be extended for database assistants by combining SQL generation with schema retrieval, RAG, SQL validation, and controlled database execution. |
|
|
| For dynamic database environments, schema information should be retrieved at inference time rather than relying only on information learned during fine-tuning. |
|
|
| Model Card Authors |
|
|
| - Author: tamilanda |
|
|
| Model Card Contact |
|
|
| For questions, issues, or contributions, please use the model repository's issue/discussion section. |
|
|
| Framework Versions |
|
|
| - PEFT: "0.17.1" |
| - Transformers: Hugging Face Transformers |
| - PyTorch: PyTorch |
