README.md

---
license: cc-by-4.0
task_categories:
- visual-question-answering
- image-classification
language:
- en
tags:
- additive-manufacturing
- laser-powder-bed-fusion
- meltpool
- near-infrared
- thermal-monitoring
- vqa
pretty_name: NIR Meltpool Thermal-State VQA
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
  - split: validation
    path: data/validation-*.parquet
  - split: test
    path: data/test-*.parquet
dataset_info:
  features:
  - name: query
    dtype: string
  - name: image
    dtype: image
  - name: annot
    dtype: string
  - name: reasoning
    dtype: string
  - name: cate
    dtype: string
  - name: task
    dtype: string
  - name: metadata
    struct:
    - name: sample
      dtype: int32
    - name: source_tif
      dtype: string
    - name: label_id
      dtype: int32
    - name: label_name
      dtype: string
    - name: time
      dtype: int64
    - name: position_X
      dtype: float64
    - name: position_Z
      dtype: float64
    - name: V
      dtype: float64
    - name: image_height
      dtype: int32
    - name: image_width
      dtype: int32
    - name: mean_intensity_8bit
      dtype: float64
    - name: bright_core_fraction
      dtype: float64
    - name: norm_low_uint16
      dtype: float64
    - name: norm_high_uint16
      dtype: float64
    - name: split
      dtype: string
---

# NIR Meltpool Thermal-State VQA

A visual-question-answering dataset for training vision-language models to classify
the **thermal state of a laser powder-bed-fusion (LPBF) meltpool** from a single
near-infrared (NIR) frame. Derived from the `hyperspectral_nir_meltpool_dataset`
(sample 588).

## Task

Given one 250x250 NIR meltpool image, answer the fixed `query` by choosing one of
six thermal-state classes.

| id | label | id | label |
|----|-------|----|-------|
| 0 | baseline | 3 | strong underheat |
| 1 | edge | 4 | overheat |
| 2 | underheat | 5 | strong overheat |

> **Classes present in this release:** only sample 588 images were available, so
> only **baseline, edge, strong underheat** appear. The `query` still lists all six
> options and the schema supports all six for when the remaining samples are added.

## Record schema

Each record has the fields:

- **query** — the question (fixed, lists all six options).
- **image** — the NIR frame (embedded PNG; HF `Image` feature).
- **annot** — ground-truth class name (the answer), from the source CSV `label`.
- **reasoning** — a grounded chain-of-thought referencing the frame's NIR intensity.
- **cate** — `meltpool_thermal_state`.
- **task** — `meltpool_thermal_state_classification`.
- **metadata** — provenance & process signals: source TIF name, `label_id`,
  stage position (`position_X`, `position_Z`), `V`, capture `time`, per-image
  intensity stats, and the global normalization constants.

## Splits

Total: **77,280** records.

| split | count | per class |
|-------|-------|-----------|
| train | 61,825 | baseline: 39,654, edge: 13,051, strong underheat: 9,120 |
| validation | 7,728 | baseline: 4,957, edge: 1,631, strong underheat: 1,140 |
| test | 7,727 | baseline: 4,956, edge: 1,631, strong underheat: 1,140 |

**Split strategy — leakage-aware.** Frames are ~2 us apart and adjacent frames are
near-duplicates. Splitting is *time-block stratified*: within each label the frames
are ordered by capture time and cut into 10 contiguous blocks (8 train / 1
validation / 1 test), so near-duplicate neighbours stay in the same split.

## Image processing

Source TIFs are single-channel `uint16`. They are mapped to 8-bit RGB PNG with a
**single global** linear normalization (clip to the dataset-wide 1st/99th
percentile, `[norm_low_uint16, norm_high_uint16]`, then scale to 0-255) and
channel-replicated. Global (not per-image) normalization is deliberate: absolute
brightness is the physical cue that distinguishes under- vs over-heated pools, and
per-image contrast stretching would destroy it.

## Usage

```python
from datasets import load_dataset
ds = load_dataset("YOUR_USERNAME/YOUR_REPO")
ex = ds["train"][0]
ex["image"]      # PIL.Image
ex["query"], ex["annot"], ex["reasoning"]
ex["metadata"]["label_id"]
```

## Provenance & license

Converted from `hyperspectral_nir_meltpool_dataset.csv` +
`hyperspectral_nir_meltpool_images_2.zip` (sample 588). Ground-truth labels come
directly from the CSV `label` column, matched to images by exact `filename`
(verified 1:1 for all 77,280 frames). Released under **CC-BY-4.0**; if the upstream
meltpool dataset carries a different license, update this field before publishing.