Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is released for research use and access is reviewed and granted manually by the maintainers. Licensing of the underlying source is NOT yet cleared (see the card). Please state your name, affiliation, and intended use.
Log in or Sign Up to review the conditions and access this dataset content.
Roles
Roles: canon repo — annot is the source label (normal / underheat / overheat / edge for the 4-class question, normal / anomaly for the binary one), kept machine-parseable as the gold for verification and reward parsing; the model reads query + image, and metadata.output_type says which of the two questions a record asks. The reasoning column is empty, so this repo is not itself a training view for chain-of-thought. metadata is provenance and must never be fed to a model.
AM12_MeltPoolKinetics
NIR melt-pool ROI frame -> thermal state: 4-way {normal,underheat,overheat,edge} (T-B2) and binary normal/anomaly (T-B1, edge excluded). Category B, task T-B2 + T-B1, in the unified Smart-Manufacturing SFT schema.
Records
81,099 records (train=81,099). Model input: a single NIR melt-pool ROI image (196x196), bytes embedded in the image column.
Unified SFT schema (7 fields)
| field | type | meaning |
|---|---|---|
query |
str | the question / instruction (model input) |
image |
Image | null | the INPUT image (bytes embedded) — null for tabular records |
annot |
str | the answer — for this dataset: the thermal-state label (text); the task column selects the 4-way vs binary framing |
reasoning |
null | no native CoT in this dataset |
cate |
"B" | SFT category |
task |
"T-B2 + T-B1" | unified task id |
metadata |
str (JSON) | split, provenance, units, input, license, and (for AM11/13) taxonomy_note |
Notes
Two tasks share one dataset (distinguished by the task column: 45,056 T-B2 + 36,043 T-B1). Only source16 carries genuine per-frame labels. The LOBO split is kept in metadata.lobo_split (not an HF split). Two misfiled non-data PNGs in the source are excluded.
Provenance & licensing
Underlying source: MeltPoolKinetics figshare compilation, source16 = NIST AMMT; DOI 10.6084/m9.figshare.28200101. Upstream license: figshare compilation — per-source terms (source16 = NIST AMMT); verify per source.
Converted read-only into the unified schema; conversion + publish scripts live in
AI4Manufacturing/forge_model
(AM12_MeltPoolKinetics/ + publish/push_am_to_hf.py). Access is gated (manual approval); clear the
upstream licence before any onward redistribution.
Query text — pooled paraphrases (v2)
This repository ships 2 question forms over the same images, and each draws from its own pool in common/vision_query_pools.json (metadata.query_template is the index within that form's pool; metadata.output_type says which form a record is):
AM12_MeltPoolKinetics/binary— 36,043 records, 37 gate-verified paraphrases (37 in use, top share 2.9%); template 0 is v1's wording byte for byte.AM12_MeltPoolKinetics/orig— 45,056 records, 37 gate-verified paraphrases (37 in use, top share 2.8%); template 0 is v1's wording byte for byte.
Template ↔ gold independence on this build: 81,099 records, 74 templates, worst template p = 0.059, alpha 1.4e-04, 0 flagged → PASS.
Frame-size floor (common/lazy_floors.py, the standing (width, height)-only row): vacuous by construction — all 81,099 images share one frame size.
Answers, images, masks, split and every other field are byte-identical to v1: this revision was issued from the published parquet itself, not rebuilt from source, and the pixel-identity guard ran on the embedded images (§8 below).
Image identity — measured at this republish
Image identity. This repository's metadata carries no file hash; metadata.pixel_sha256 (added at this republish, one per image; pixel_sha256_images per entry of images) hashes the decoded image (mode | size | pixels) and settles whether an image is duplicated.
Measured at build time, not asserted afterwards — a violation aborts the build and names the offending records:
| images checked | 81,099 |
| distinct by decoded pixels | 45,056 |
| images carrying more than one record | 36,043 |
| images on both sides of the split | 0 |
⚠ This dataset declares a many image-identity policy, so the row above is expected to be non-zero: BY DESIGN: this repository asks TWO questions about each melt-pool frame — the 4-class state task (45,056 records) and the binary normal/anomaly task (36,043 records, the scan-boundary class excluded) — so 36,043 frames carry two records, one per question, with different queries and different answers. No frame is duplicated within a question, and the 9,013 edge frames appear only in the 4-class task. metadata.output_type says which question a record asks Images are still forbidden from crossing the split, and none do.
- Downloads last month
- 63