emirkaanozdemr/cukurova-3b-instruct

Cukurova-3B Reasoning Model

Overview

Cukurova-3B Instruct is a domain-adapted large language model designed for structured reasoning and decision generation from real-world data.

The model is capable of:

• Interpreting structured sensor inputs
• Performing multi-step reasoning
• Producing structured decision outputs
• Simulating autonomous control agents

The architecture allows the model to function as a general reasoning engine for data-driven decision systems.

Although the model is demonstrated using agricultural sensor data, the reasoning framework can be applied to any structured environment monitoring or automation system.

---

Base Model

This model is fine-tuned from:

Llama 3.2 3B Instruct

The base model provides strong instruction-following capabilities, while the fine-tuning process teaches the model how to:

• reason over structured data
• explain decisions
• generate machine-readable actions

---

Training Dataset

Fine-tuning was performed using the Cotton Field Dataset, which provides structured sensor environments paired with reasoning explanations and action outputs.

Dataset link:

https://huggingface.co/datasets/emirkaanozdemr/cotton-field-dataset

The dataset simulates autonomous control scenarios where an AI system must interpret environmental data and decide appropriate system actions.

While the dataset focuses on agricultural environments, the training objective teaches the model a general pattern of reasoning from state → explanation → action.

---

Dataset Structure

Each dataset sample contains three key components.

1. Environment State (Input)

Structured sensor measurements describing the current system state.

Example:

{
  "soil_moisture": 47.84,
  "air_humidity": 71.55,
  "sunlight": 410.96,
  "plant_stage": "FLOWERING",
  "rain_forecast": 74.74
}

These structured inputs simulate real-world monitoring systems, such as IoT sensor networks.

---

2. Reasoning Target

A natural language explanation describing why certain decisions should be made based on the system state.

Example:

Rain forecast is high so irrigation is disabled to avoid waterlogging.
UV light is enabled to supplement sunlight during flowering stage.

This step encourages the model to produce interpretable reasoning traces.

---

3. Action Output

A structured control decision generated from the reasoning process.

Example:

{
  "irrigation_on": false,
  "water_liters": 0,
  "pesticide_on": false,
  "uv_light_on": true,
  "uv_intensity": 39.6,
  "shade_percent": 0,
  "fan_speed": "MEDIUM"
}

Structured outputs allow the model to integrate directly with automated control systems or APIs.

---

Model Objective

The model learns a general reasoning pipeline:

Structured Data
     ↓
Reasoning Process
     ↓
Structured Decision Output

This enables the model to act as a reasoning engine for autonomous systems.

---

Example Prompt

Sensor data:

Soil moisture: 47
Air humidity: 71
Sunlight: 410
Plant stage: FLOWERING
Rain forecast: 74
Disease risk: 0.27

Explain the reasoning and decide the correct actions.

---

Example Output

Reasoning:
Rain forecast is high so irrigation is disabled.
Disease risk is low so pesticide is not applied.
Additional UV light is used during flowering stage.

Actions:
{
 "irrigation_on": false,
 "water_liters": 0,
 "pesticide_on": false,
 "uv_light_on": true
}

---

Use Cases

Potential applications include:

• Autonomous system controllers
• IoT decision systems
• Environmental monitoring AI
• Smart agriculture
• Industrial automation assistants
• Research on reasoning-based AI agents

---

Training Method

Fine-tuning was performed using instruction tuning with reasoning targets.

Training signals include:

• Structured state → reasoning generation
• Reasoning → structured action prediction

This setup encourages the model to learn explainable decision making rather than direct prediction.

---

Citation

If you use this model or dataset in research, please cite:

Cukurova-3B Instruct – Emir Kaan Özdemir
Cotton Field Dataset – Emir Kaan Özdemir

---

Author

Emir Kaan Özdemir