README.md

metadata

license: apache-2.0
task_categories:
  - visual-question-answering
  - object-detection
tags:
  - multimodal
  - RL
  - vision-agent
  - tool-learning
configs:
  - config_name: default
    data_files:
      - split: train
        path: reasoning_rl.parquet
    features:
      - name: data_source
        dtype: string
      - name: prompt
        list:
          - name: content
            dtype: string
          - name: role
            dtype: string
      - name: images
        list:
          - name: bytes
            dtype: binary
          - name: name
            dtype: string
          - name: path
            dtype: string
      - name: ability
        dtype: string
      - name: env_name
        dtype: string
      - name: reward_model
        dtype: string
      - name: extra_info
        dtype: string
      - name: agent_name
        dtype: string

---

APEX-GRO-RL Dataset

1. Introduction

APEX-GRO-RL is a multimodal dataset specifically curated for training Visual Analysis Agents using Reinforcement Learning (RL). It integrates visual counting and visual grounding tasks, designed to teach agents how to autonomously plan reasoning behaviors and invoke active perception tools (such as zoom_in) to inspect dense or small targets in high-resolution images.

The data format seamlessly fits training environments like visual_toolbox, where system observations and structured tool-call formatting are required.

2. Dataset Structure

The dataset is stored in Apache Parquet format. Each entry contains the following fields:

Field Name	Type	Description
data_source	string	Source of the original data (APEX-GRO).
prompt	list	Multi-turn style conversational prompt template containing system guidelines and the formatted user question.
images	list	List of images related to the sample. Each image dict contains name, path, and raw image binary data encoded in WebP format.
ability	string	Task capability type: counting or grounding.
env_name	string	Target environment name for RL setup (visual_toolbox).
reward_model	string (JSON)	Configuration for reward calculation, including ground_truth and matching style.
extra_info	string (JSON)	Metadata tracking including original dataset index, original resolution, and target relative bounding boxes (rel_bboxs).
agent_name	string	Target agent architecture type (tool_agent).