Datasets:

findingdory
/

findingdory-video-sft

---
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).

	---
	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).