Agri-CM3-Vision-Unsloth

An English vision-only dataset prepared for fine-tuning Vision Language Models (VLMs) with Unsloth.

This is a reformatted subset of the original HIT-Kwoo/Agri-CM3 benchmark — a large-scale Chinese agricultural pest and disease dataset. We extracted only the English vision splits, keeping all image-based tasks and formatting them in the ShareGPT conversation format compatible with Unsloth fine-tuning.

Purpose

This dataset was specifically prepared for Vision Language Model (VLM) fine-tuning using Unsloth. The goal is to train models that can:

• Identify crops, diseases, and pests from photos
• Reason about visible symptoms
• Apply agricultural domain knowledge from visual input

Dataset Structure

Subset	Task	Train	Validation
Disease-P	Identify crop/disease from photo	~1,183	100
Disease-P-M	Identify + reason about symptoms	~1,183	100
Disease-P-M-K	Identify + reason + apply knowledge	~1,183	100
Pest-P	Identify pest from photo	~1,183	100
Pest-P-K	Identify pest + apply knowledge	~1,183	100
all-image	All 5 subsets merged	~8,939	~500

Columns

• image — high-resolution crop/pest photo (~1800×1500px average)
• messages — conversation in ChatML format (role: user / role: assistant)

Usage in Unsloth Studio

1. Dataset: farukalamai/Agri-CM3-Vision-Unsloth
2. Subset: pick any (recommended: all-image for full training, Disease-P-M-K for disease focus)
3. Train Split: train
4. Evaluation Split: validation
5. Target Format: ChatML

Usage in Notebook (Colab / Local)

from datasets import load_dataset

dataset = load_dataset("farukalamai/Agri-CM3-Vision-Unsloth", "all", split="train")

def to_unsloth(sample):
    return {"messages": [
        {"role": "user", "content": [
            {"type": "image", "image": sample["image"]},
            {"type": "text",  "text": sample["messages"][0]["content"][0]["text"]},
        ]},
        {"role": "assistant", "content": [
            {"type": "text", "text": sample["messages"][1]["content"][0]["text"]},
        ]},
    ]}

converted = [to_unsloth(s) for s in dataset]

Original Dataset

This dataset is derived from:

Agri-CM3: A Chinese Massive Multi-modal, Multi-level Benchmark for Agricultural Understanding and Reasoning Wang et al., ACL 2025 HIT-Kwoo/Agri-CM3

Original license: Apache 2.0

Citation

@inproceedings{wang2025agricm3,
  title={Agri-CM3: A Chinese Massive Multi-modal, Multi-level Benchmark for Agricultural Understanding and Reasoning},
  author={Wang et al.},
  booktitle={ACL},
  year={2025}
}