configs:
- config_name: default
data_files:
- split: train
path: '*/train/*.arrow'
- split: val
path: '*/val/*.arrow'
- split: test
path: '*/test/*.arrow'
- config_name: seq_len_1
data_files:
- split: train
path: seq_len_1/train/*.arrow
- split: val
path: seq_len_1/val/*.arrow
- split: test
path: seq_len_1/test/*.arrow
- config_name: seq_len_2
data_files:
- split: train
path: seq_len_2/train/*.arrow
- split: val
path: seq_len_2/val/*.arrow
- split: test
path: seq_len_2/test/*.arrow
- config_name: seq_len_4
data_files:
- split: train
path: seq_len_4/train/*.arrow
- split: val
path: seq_len_4/val/*.arrow
- split: test
path: seq_len_4/test/*.arrow
- config_name: seq_len_8
data_files:
- split: train
path: seq_len_8/train/*.arrow
- split: val
path: seq_len_8/val/*.arrow
- split: test
path: seq_len_8/test/*.arrow
- config_name: seq_len_16
data_files:
- split: train
path: seq_len_16/train/*.arrow
- split: val
path: seq_len_16/val/*.arrow
- split: test
path: seq_len_16/test/*.arrow
- config_name: seq_len_32
data_files:
- split: test
path: seq_len_32/test/*.arrow
- config_name: seq_len_64
data_files:
- split: test
path: seq_len_64/test/*.arrow
- config_name: seq_len_128
data_files:
- split: test
path: seq_len_128/test/*.arrow
MMReD: A Cross-Modal Benchmark for Dense Reasoning
This is the dataset & benchmark accompanying MMReD paper. It was obtained by running generation script in MMReD repository:
python scripts/generate_dataset.py
It only contains textual split of the dataset to save space as images can be deterministically generated from jsons.
To run full evaluation, training, or generate images from jsons for LVLMs, please refer to the repository.
Multi-Modal Controllable Environment = Dense Factual Haystack
MMReD introduces a concept of dense context, aiming at controllable generative evaluation of reasoning over arbitrarily long factually dense scenarios.
It contains 8 splits corresponding to different sequence lengths in environment transitions: [1, 2, 4, ..., 128]. Each split contains 24 evaluation questions with 200/50/50 train/val/test samples per question type.
Different splits grouped by sequence length are available both form evaluation & training, with training samples generated up to sequence length of 16.
from datasets import load_dataset
train_16 = load_dataset("dondo-sss/mmred", "seq_len_128")["train"]
test_128 = load_dataset("dondo-sss/mmred", "seq_len_128")["test"]
Questions are generally divided into 2 groups - resembling standard NIAH evaluation & our introduced Dense Long Context evaluation:
| ID | Question template | Dataset Name |
|---|---|---|
| NIAH | ||
| FA-FA-R | In which room did [C] first appear? | room_on_char_first_app |
| FA-CCFA-R | In which room was [C1] when [C2] first appeared in the [R]? | char_on_char_first_app |
| FA-FR-C | Who was the first to appear in the [R]? | first_at_room |
| FA-RCFA-C | Who was in the [R1] when [C] first appeared in the [R2]? | char_at_frame |
| FA-NRFA-I | How many characters were in the [R1] when [C] first appeared in the [R2]? | n_room_on_char_first_app |
| FI-FA-R | In which room was [C] at the final step? | final_app |
| FI-CCFA-R | In which room was [C1] when [C2] made their final appearance in the [R]? | char_on_char_final_app |
| FI-LR-C | Who was the last to appear in the [R]? | last_at_room |
| FI-RCFA-C | Who was in the [R1] when [C] made their final appearance in the [R2]? | char_at_frame |
| FI-NRFA-I | How many chars were in the [R1] when [C] made their final app in the [R2]? | n_room_on_char_final_app |
| FX-CF-R | In which room was [C] at step [X]? | room_at_frame |
| FX-RF-C | Who was in the [R] at step [X]? | char_at_frame |
| FX-CCF-C | Who was in the same room as [C] at step [X]? | char_on_char_at_frame |
| FX-NCF-I | How many other characters were in the same room as [C] at step [X]? | n_char_at_frame |
| FX-NE-I | How many rooms were empty at step [X]? | n_empty |
| LC | ||
| LC-RE-R | Which room was empty for the [comp] steps? | room_empty |
| LC-WS-R | In which room did [C] spend the [comp] time? | where_spend |
| LC-CR-R | Which room was crowded (three or more people) for the most steps? | crowded_room |
| LC-WHS-C | Who spent the [comp] time in the [R]? | who_spend |
| LC-SA-C | Who spent the [comp] time alone in the rooms? | spend_alone |
| LC-ST-C | With whom did [C] spend the [comp] time together in the same room? | spend_together |
| LC-SR-I | How many steps did [C] spend in the [R]? | steps_in_room |
| LC-RV-I | How many different rooms did [C] visit? | rooms_visited |
| LC-CC-I | How many times did a crowd (three or more people in one room) appear? | crowd_count |