This model, Llama-3.2-3B-Instruct-Spatial-SQL-1.0, is an 3B, narrow use case, text to spatial SQL, lightly fine-tuned model. In general, its primary use case is the Natural Language command adaptation of particular geographic spatial functions as normally defined in pure SQL. Data input should be a combination of an English prefix in the form of a question, and a coordinate prompt injection, likely from an active mapping system application coordinate list. Output is PostGIS spatial SQL.

There are four primary geographic functions released in version 1.0.

Model developer: Mark Rodrigo

Github: https://github.com/mprodrigo/spatialsql

Model Architecture: The model is a QLoRA / Supervised Fine Tuning (SFT)

Model Input / Output Overview:

Input: Text plus coordinate prompt injection.
Output: PostGIS spatial SQL
NOTE: Inputs and outputs are in meters and or geographic decimal degrees WGS 84 coordinates.

Function	Question Input	Geo Input	SQL Execution Output
Area	Area question	Polygon	Number - Area sq meters
Centroid	Center question	Polygon	Point
Buffer	Buffer distance	Point	Polygon
Length	Length question	Line	Number - Length in meters

Example Prompt / Prompt File

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

You are a helpful assistant. You are an expert at PostGIS and Postgresql and SQL and psql.

<|eot_id|><|start_header_id|>user<|end_header_id|>

### Instruction: Write a PostGIS SQL statement for the following.

### Input:

{input}

### Response:

<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Examples

AREA

### Input: What is the area for the polygon? : 'Polygon ((-3.7515154 40.3855551, -3.7514972 40.3856581, -3.7507005 40.3855767, -3.7507167 40.3854722, -3.7515154 40.3855551))'

### Response: SELECT ST_Area(geog) As area FROM (select 'Polygon ((-3.7515154 40.3855551, -3.7514972 40.3856581, -3.7507005 40.3855767, -3.7507167 40.3854722, -3.7515154 40.3855551))' :: geography geog) subquery;

CENTROID

### Input: What is the centroid for the polygon? : 'Polygon ((-3.6934636 40.4808785, -3.6933352 40.4811486, -3.6930125 40.4810598, -3.693141 40.4807897, -3.6934636 40.4808785))'

### Response: SELECT ST_AsText(ST_Centroid(geog)) As centroid FROM (select 'Polygon ((-3.6934636 40.4808785, -3.6933352 40.4811486, -3.6930125 40.4810598, -3.693141 40.4807897, -3.6934636 40.4808785))' :: geography geog) subquery;

BUFFER

### Input: What is the thousand meter buffer for the following point? : 'Point(-8.7522658 41.3862664)'

### Response: SELECT ST_AsText(ST_Buffer(geog, 1000)) as buffer FROM (select 'Point(-8.7522658 41.3862664)' :: geography geog) subquery;

LENGTH

### Input: How long is the line? : 'LINESTRING (-3.6976693 40.4263178, -3.6986082 40.4258729)'

### Response: SELECT ST_Length(geog) As length FROM (select 'LINESTRING (-3.6976693 40.4263178, -3.6986082 40.4258729)' :: geography geog) subquery;

A Few Known Question Variation Examples

AREA

What is the area for the geometry?

What is the area for this polygon?

CENTROID

What is the centroid for the geometry?

What is the center point of the polygon?

BUFFER

What is the 100 meter buffer for the following point?

Buffer the following point a thousand meters.

What is the 1000 meter buffer for the following point?

LENGTH

What is the length of the line?

How long is this line?

llama.cpp / Hyperparameter Recommendations For Inference

max context ~ 8,000 or lower

top k ~ 100

temp ~ .4-.5 or lower

Agent Considerations

Agents are being considered as a separate project. Agents would mostly be related to pulling the coordinates from a mapping UI, and executing the SQL from responses against a PostGIS database.

Further Reference - link this

https://postgis.net/docs/manual-3.3/PostGIS_Special_Functions_Index.html#PostGIS_GeographyFunctions

Evaluation data

More information needed

Training data

Custom synthetic

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 3e-05
train_batch_size: 10
eval_batch_size: 3
distributed_type: multi-GPU
num_devices: 2
optimizer: Adam 8bit
lr_scheduler_type: linear
num_epochs: 5

Training results

Training Loss	Step	Validation Loss
1.3788	10	1.3543
0.8504	20	0.8671
0.5990	30	0.6501
0.5836	40	0.5885
0.5713	50	0.5607
0.5481	60	0.5434
0.5221	70	0.5342
0.4983	80	0.5259
0.4914	90	0.5219
0.4754	100	0.5205

Framework versions

Transformers 4.45.1
Pytorch 2.4.1
peft 0.13.0
Datasets 3.0.1
Tokenizers 0.20.0

markrodrigo/Llama-3.2-3B-Instruct-Spatial-SQL-1.0

Model Information