MaolinLuo commited on
Commit
2ad8a8d
·
1 Parent(s): 0a4545c

Update MRCL

Browse files
MedBookVQA/jsons/test/data.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:d311ae232df92988ca790878d1f3d651beb1a14acac01b0baac812718acdb680
3
- size 526563
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0d37f1bb78610eb13cfca793f9239260f4602b694c87cb39febde368c659c617
3
+ size 525974
MedBookVQA/jsons/train/data.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5eecd65391362320f814c6c4c591007a24a1873c5b37eadfc96910b469b1466e
3
- size 1240456
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:00780f8efa29271ebe39edf2ec709ea95b1702dbb7071a5c4d65f93d63e46f1c
3
+ size 1239082
README.md CHANGED
@@ -45,6 +45,14 @@ configs:
45
  path: FinMME/jsons/test/data.json
46
  ---
47
 
 
 
 
 
 
 
 
 
48
  # MRCL: Multimodal Reasoning Continual Learning
49
 
50
  MRCL is a five-stage benchmark for studying catastrophic forgetting during continual post-training of vision-language models. It brings together recent, challenging, and reasoning-intensive multimodal datasets spanning medical understanding, navigation and planning, geometry, visual-spatial reasoning, and financial chart analysis.
@@ -71,7 +79,7 @@ Each stage provides a training split for adaptation and a held-out test split fo
71
  | ---: | --- | --- | ---: | ---: | --- |
72
  | 1 | MedBookVQA | Medical visual understanding | ~1.4K | ~0.6K | Multiple choice |
73
  | 2 | Navigation | Visual navigation and planning | ~1.5K | ~0.5K | Direction sequence |
74
- | 3 | We-Math2.0 | Multimodal mathematical reasoning | ~5.7K | ~3.7K | Numerical and multiple choice |
75
  | 4 | Puzzle | Visual-spatial reasoning | ~6.0K | ~0.4K | Multiple choice |
76
  | 5 | FinMME | Financial chart understanding | ~8.0K | ~1.0K | Numerical and multiple-answer choice |
77
 
 
45
  path: FinMME/jsons/test/data.json
46
  ---
47
 
48
+ # Important Notice: Dataset Update (2026-10-12)
49
+
50
+ We identified and fixed an answer leakage issue caused by inconsistent option spacing in **MedBookVQA**. This specific issue is limited to MedBookVQA and does not involve the other four MRCL task datasets.
51
+
52
+ Additionally, we removed duplicate or empty samples from We-Math2.0, affecting approximately 7% of the training set and 0.1% of the test set.
53
+
54
+ **The current dataset has been corrected and is ready for use.**
55
+
56
  # MRCL: Multimodal Reasoning Continual Learning
57
 
58
  MRCL is a five-stage benchmark for studying catastrophic forgetting during continual post-training of vision-language models. It brings together recent, challenging, and reasoning-intensive multimodal datasets spanning medical understanding, navigation and planning, geometry, visual-spatial reasoning, and financial chart analysis.
 
79
  | ---: | --- | --- | ---: | ---: | --- |
80
  | 1 | MedBookVQA | Medical visual understanding | ~1.4K | ~0.6K | Multiple choice |
81
  | 2 | Navigation | Visual navigation and planning | ~1.5K | ~0.5K | Direction sequence |
82
+ | 3 | We-Math2.0 | Multimodal mathematical reasoning | ~5.3K | ~3.7K | Numerical and multiple choice |
83
  | 4 | Puzzle | Visual-spatial reasoning | ~6.0K | ~0.4K | Multiple choice |
84
  | 5 | FinMME | Financial chart understanding | ~8.0K | ~1.0K | Numerical and multiple-answer choice |
85
 
We-Math2/jsons/test/data.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:686b43e6493afabd2134ae9b4d7c7d5d247a659c8348f97642a307ae61a86682
3
- size 2864772
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:1f89c423989acb55589b761adf55b1fe1b9072c387b4eca6b6b47f7a157d2fc9
3
+ size 2861317
We-Math2/jsons/train/data.json CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1982122dd597dada8dc0479ef9a116fcaaaf9d985d0fb7e6569dc76a5c6b8484
3
- size 3527727
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e801e940fc72f88d9a6b6a6bb29b130f5af9ec534d13f1005bb06aebd92ad79d
3
+ size 3296334