jackcwf/DeepSeek-R1-Distill-Llama-4bit-sql-model

About Model

Fine-tuning is used to convert SQL language into natural language, making it easier for users to understand the business meaning of SQL queries. This fine-tuned model is based on the unsloth framework AND uses the DeepSeek-R1-Distill-Llama-8B pre-trained model under unsloth.

DataSet

b-mc2/sql-create-context

Model train

1. train/loss: This chart shows the model's loss during training. As the training steps (global step) increase, the loss value drops sharply 和 then stabilizes, indicating that the model is gradually converging.
2. train/learning_rate: This chart shows how the learning rate changes over training steps. From the chart, we can see that the learning rate decreases as training progresses, which is likely part of a learning rate decay strategy to prevent the model from oscillating in the later stages of training.
3. train/grad_norm: This chart displays the change in gradient norm over training steps. The decrease in gradient norm suggests that the gradients are stabilizing, reducing instability during training.
4. train/global_step: This chart shows the increase in global training steps. As the training progresses, the step count gradually increases, indicating the progress of the training process.
5. train/epoch: This chart represents the progress of each training epoch. As the global steps increase, the epoch count also steadily grows.

Inference results before 和 after model training:

Prompt

Define SQL query for testing

This is a complex customer analysis query used to test the understanding of the model

query1 = """

SELECT
    pc.category_name,
    p.product_name,
    COUNT(DISTINCT o.customer_id) AS unique_customers,
    COUNT(oi.order_id) AS total_sales,
    SUM(oi.quantity) AS total_quantity_sold,
    ROUND(AVG(oi.unit_price), 2) AS avg_selling_price,
    SUM(oi.quantity * oi.unit_price) AS total_revenue,
    ROUND(SUM(oi.quantity * oi.unit_price) / COUNT(DISTINCT o.customer_id), 2) AS revenue_per_customer,
    MAX(o.order_date) AS last_sale_date,
    MIN(o.order_date) AS first_sale_date
FROM product_categories pc
JOIN products p ON pc.category_id = p.category_id
JOIN order_items oi ON p.product_id = oi.product_id
JOIN orders o ON oi.order_id = o.order_id
WHERE
    o.order_date >= '2024-01-01'
    AND o.order_status = 'completed'
GROUP BY
    pc.category_name,
    p.product_name
HAVING
    total_revenue > 10000
ORDER BY
    total_revenue DESC,
    unique_customers DESC
LIMIT 15;

Explain use case of this query. """

CREATE TABLE product_categories (category_name VARCHAR, product_name VARCHAR, customer_id INT, order_id INT, order_date DATETIME, status VARCHAR) FROM product_categories JOIN products JOIN order_items JOIN orders WHERE order_date >= '2024-01-01' AND order_status = 'completed' GROUP BY category_name, product_name HAVING total_revenue > 10000

This query analyzes the relationship between product categories and customer orders to identify categories and products with the highest revenue in completed orders since 2024. It can help a business understand which products are generating significant revenue and which products are performing well in the category to inform inventory planning and pricing strategies. Below is the business use case for the query:

Use Case:

This query analyzes the relationship between product categories and customer orders to identify categories and products with the highest revenue in completed orders since 2024. It can help a business understand which products are generating significant revenue in the category and which products are performing well in the category to inform inventory planning and pricing strategies.<｜end▁of▁sentence｜>

Model Download

Model	Base Model	下载
unsloth/DeepSeek-R1-Distill-Llama-8B	DeepSeek-R1-Distill-Llama-8B	🤗 HuggingFace

Usage

If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files.

Uploaded model

• Developed by: datalabs-ai
• License: apache-2.0
• Finetuned from model : unsloth/DeepSeek-R1-Distill-Llama-8B

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.