devpatel1012 commited on
Commit
83ee74f
·
verified ·
1 Parent(s): 3324c9a

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +158 -15
README.md CHANGED
@@ -1,26 +1,169 @@
1
  ---
 
2
  license: apache-2.0
 
 
3
  task_categories:
4
- - question-answering
5
- - image-to-text
6
- language:
7
- - en
8
  tags:
9
- - vlm
10
- - visual-question-answering
11
- - dataset-polishing
 
 
 
 
12
  size_categories:
13
- - 100K<n<1M
 
 
 
 
 
14
  ---
15
 
16
- # Polished VLM Dataset
17
 
18
- This dataset contains polished and augmented Visual Question Answering (VQA) pairs designed for training and evaluation of Vision-Language Models (VLMs).
 
 
 
 
19
 
20
- ## Dataset Structure
 
 
 
 
 
21
 
22
- - **Columns:** `question`, `answer`, `image`, `original_row_index`, `original_filename`, `original_caption`
23
- - **Total Rows:** 254,138
 
24
 
25
- ## Summary
26
- This dataset is uploaded directly from cleaned, parquet-sharded preprocessing outputs.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
  ---
2
+ pretty_name: MedOptm Surgical VQA (polished)
3
  license: apache-2.0
4
+ language:
5
+ - en
6
  task_categories:
7
+ - visual-question-answering
8
+ - image-to-text
 
 
9
  tags:
10
+ - medical
11
+ - biomedical
12
+ - surgical
13
+ - vqa
14
+ - visual-question-answering
15
+ - multimodal
16
+ - biomedica
17
  size_categories:
18
+ - 100K<n<1M
19
+ configs:
20
+ - config_name: default
21
+ data_files:
22
+ - split: train
23
+ path: data/train-*.parquet
24
  ---
25
 
26
+ # MedOptm Surgical VQA (polished)
27
 
28
+ A biomedical **visual-question-answering (VQA)** dataset: each example is a
29
+ medical/surgical figure paired with a natural-language **question** and a short
30
+ **answer**. It was used to train
31
+ [**MedOptm**](https://github.com/Devpatel1012/MedOptm-VLM), a compact medical
32
+ vision-language model (CvT / CLIP-ViT + Llama-3.2-1B).
33
 
34
+ > **This VQA dataset is derived from the **BIOMEDICA** vision–caption data.**
35
+ > The questions and answers were **generated from BIOMEDICA figure captions** —
36
+ > the images and their captions come from BIOMEDICA (itself built from PubMed
37
+ > Central Open Access), and each `(question, answer)` pair was derived from the
38
+ > figure's caption. The original caption is kept in every row (`original_caption`)
39
+ > so the derivation stays auditable.
40
 
41
+ > ⚠️ **Research use only — not for clinical use.** Answers are derived from figure
42
+ > captions, not written or reviewed by clinicians, and can be incomplete or
43
+ > wrong. Do not use this dataset to make or support medical decisions.
44
 
45
+ ---
46
+
47
+ ## Where the code lives
48
+
49
+ Everything that loads, cleans, inspects and consumes this dataset is open-source
50
+ in the MedOptm repository:
51
+
52
+ - **GitHub repo:** <https://github.com/Devpatel1012/MedOptm-VLM>
53
+ - **Data pipeline code:** [`src/medoptm/data/`](https://github.com/Devpatel1012/MedOptm-VLM/tree/main/src/medoptm/data)
54
+ - [`dataset.py`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/src/medoptm/data/dataset.py) — loads these parquet files, decodes the images, and builds the train/val/test DataLoaders.
55
+ - [`preprocessing.py`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/src/medoptm/data/preprocessing.py) — the `light_scrub()` text cleaner that removed demographics/measurements from the answer text.
56
+ - [`sampling.py`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/src/medoptm/data/sampling.py) — pulls small random samples out for inspection.
57
+ - **BIOMEDICA → VQA inspection notebooks:**
58
+ [`01_biomedica_data_inspection.ipynb`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/notebooks/01_biomedica_data_inspection.ipynb)
59
+ and [`02_biomedica_vqa_inspection.ipynb`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/notebooks/02_biomedica_vqa_inspection.ipynb)
60
+ show the BIOMEDICA caption data and the derived VQA side by side.
61
+
62
+ ---
63
+
64
+ ## Dataset structure
65
+
66
+ A single `train` split, stored as parquet shards under `data/`. Each row:
67
+
68
+ | Column | Type | Description |
69
+ |---|---|---|
70
+ | `image` | image (bytes) | The biomedical/surgical figure from BIOMEDICA. With 🤗 `datasets` this is cast to a PIL image automatically. |
71
+ | `question` | string | A question about the image. ~56 characters on average. |
72
+ | `answer` | string | The answer, derived from the figure caption. ~77 characters on average. |
73
+ | `original_caption` | string | The BIOMEDICA figure caption the Q&A was derived from. ~219 characters on average. Kept for provenance and for caption-style training. |
74
+ | `original_filename` | string | The BIOMEDICA parquet shard this row came from. |
75
+ | `original_row_index` | int64 | The row index within that shard (`-1` where unknown). |
76
+
77
+ The dataset is **not pre-split** — MedOptm splits it at load time (80/10/10
78
+ train/val/test, `seed=42`).
79
+
80
+ ### Example
81
+
82
+ ```text
83
+ question: "Where in the spine were pedicle screws placed?"
84
+ answer: "The pedicle screws were placed in the lumbar spine."
85
+ original_caption: "The pedicle screws were placed percutaneously (A). The positions of
86
+ the screws are shown in the X-ray lateral view (B), and the
87
+ intervertebral ..."
88
+ ```
89
+
90
+ ---
91
+
92
+ ## Usage
93
+
94
+ ```python
95
+ from datasets import load_dataset
96
+
97
+ ds = load_dataset("devpatel1012/polished-vlm-dataset", split="train")
98
+
99
+ row = ds[0]
100
+ print(row["question"], "->", row["answer"])
101
+ row["image"] # a PIL image
102
+ ```
103
+
104
+ Reproduce MedOptm's exact split:
105
+
106
+ ```python
107
+ split = ds.shuffle(seed=42).train_test_split(test_size=0.2)
108
+ train = split["train"]
109
+ val_test = split["test"].train_test_split(test_size=0.5)
110
+ val, test = val_test["train"], val_test["test"]
111
+ ```
112
+
113
+ ---
114
+
115
+ ## How it was built
116
+
117
+ 1. **Source.** Figures and their captions from **BIOMEDICA** (derived from
118
+ PubMed Central Open Access).
119
+ 2. **Q&A generation.** For each figure, a `(question, answer)` pair was generated
120
+ from its BIOMEDICA caption. The caption is retained as `original_caption` so
121
+ each answer can be traced back to its source text.
122
+ 3. **Light scrubbing.** Answer text was passed through `light_scrub()`
123
+ ([`preprocessing.py`](https://github.com/Devpatel1012/MedOptm-VLM/blob/main/src/medoptm/data/preprocessing.py))
124
+ to remove demographic details and measurements.
125
+
126
+ ---
127
+
128
+ ## Provenance & licensing
129
+
130
+ This is a **derived** dataset. Its images and captions come from **BIOMEDICA**,
131
+ which is built from **PubMed Central Open Access (PMC-OA)**. PMC-OA is a *mixture*
132
+ of Creative Commons licences — some articles allow redistribution with
133
+ attribution, others carry non-commercial or no-derivatives terms. **A derived
134
+ dataset cannot be more permissive than its source.**
135
+
136
+ Before making this public:
137
+
138
+ - Read BIOMEDICA's dataset terms and licence, and PMC-OA's reuse terms.
139
+ - Set the `license:` field in the metadata block above to match.
140
+ - Keep the attribution to BIOMEDICA and PMC-OA intact.
141
+
142
+ ## Limitations
143
+
144
+ - Answers are **caption-derived**, not clinician-authored or clinically validated.
145
+ - Figures come from published literature, so they skew toward notable/illustrative
146
+ cases rather than a representative clinical distribution.
147
+ - Multi-panel figures (A/B/C …) are kept whole, so a question may refer to only
148
+ one panel.
149
+ - De-identification of the image **pixels** has not been exhaustively verified —
150
+ published figures are normally de-identified, but spot-check before relying on it.
151
+
152
+ ## Citation
153
+
154
+ Please cite **BIOMEDICA** (the upstream source) and this derived release.
155
+
156
+ ```bibtex
157
+ @misc{medoptm_surgical_vqa,
158
+ title = {MedOptm Surgical VQA (polished)},
159
+ author = {Dev Patel},
160
+ year = {2026},
161
+ url = {https://huggingface.co/datasets/devpatel1012/polished-vlm-dataset},
162
+ note = {Derived from the BIOMEDICA vision--caption dataset (PubMed Central Open Access).}
163
+ }
164
+ ```
165
+
166
+ ## Links
167
+
168
+ - 🧠 **Model & code:** <https://github.com/Devpatel1012/MedOptm-VLM>
169
+ - 🗂️ **Upstream data:** BIOMEDICA