---
dataset_info:
  features:
  - name: ep_id
    dtype: string
  - name: video
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: task_id
    dtype: string
  - name: high_level_category
    dtype: string
  - name: low_level_category
    dtype: string
  - name: num_interactions
    dtype: int64
  splits:
  - name: train
    num_bytes: 107506980
    num_examples: 79213
  - name: validation
    num_bytes: 9653447
    num_examples: 5870
  download_size: 14758637
  dataset_size: 117160427
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: validation
    path: data/validation-*
---
<center><h1>2026 ICLR Submission</h1>
</center>

Current vision-language models (VLMs) struggle with long-term memory in embodied tasks. To address this, we introduce **FindingDory**, a benchmark in Habitat that evaluates memory-based reasoning across 60 long-horizon tasks. 

In this repo, we release the FindingDory Video Dataset. Each video contains images collected from a robot’s egocentric view as it navigates realistic indoor environments and interacts with objects. This dataset was used to train and evaluate the high-level agent SFT agent in the FindingDory benchmark.

# Usage
```
from datasets import load_dataset
dataset = load_dataset("findingdory/findingdory-video-sft")
```

# Dataset Structure

| Field name                 | Description                                                                                                   |
| ------------------------- | ------------------------------------------------------------------------------------------------------------- |
| **ep\_id**                  | Episode id.                                                                                                                                                   |
| **video**                    | Relative path of the video clip.                                                                                                           |
| **question**             | Question posed to the agent based on the episode.                                                                |
| **answer**                | Ground-truth answer stored as a list of image indices                                                             |
| **task\_id**              | Identifier indicating which task template the episode belongs to (string).                     |
| **high\_level\_category** | Higl-task task category label. (Options: Single-Goal Spatial Tasks, Single-Goal Temporal Tasks, Multi-Goal Tasks). |
| **low\_level\_category**   | Fine-grained task category label. (Example categories: Interaction-Order, Room Visitation, etc)    |
| **num\_interactions**        | Number of objects the robot interacts with, during the experience collection. |
Notes:
* The validation split contains 60 tasks . The training split only contains 55 task because the 5 “Object Attributes” tasks are withheld from the training set.
* A subsampled version of the dataset (96 frames per episode) is available [here](https://huggingface.co/datasets/findingdory/findingdory-video-sft-96).