AM13_MeltpoolNet / README.md
XMQ168's picture
AM13_MeltpoolNet: T-AM2 unified SFT (viewer-friendly row groups)
21069d9 verified
|
Raw
History Blame Contribute Delete
3.98 kB
metadata
tags:
  - smart-manufacturing
  - additive-manufacturing
  - sft
license: other
pretty_name: AM13_MeltpoolNet
extra_gated_fields:
  Name: text
  Affiliation: text
  Intended use: text
extra_gated_prompt: >-
  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.

AM13_MeltpoolNet

Unit-normalized process/material inputs -> source-native 5-way melt-pool mode classification. Tabular. Category C, task T-AM2, in the unified Smart-Manufacturing SFT schema.

Records

1,236 records (train=1,236). Model input: tabular — no image.

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: a text class label, one of keyhole, LOF, balling, desirable, spatter formation
reasoning null no native CoT in this dataset
cate "C" SFT category
task "T-AM2" unified task id
metadata str (JSON) split, provenance, units, input, license, and (for AM11/13) taxonomy_note

Notes

Leakage-guarded by normalized input fields: answer/outcome columns (meltpool shape, non-mode measurements, porosity, relative density, spatter, d/l·d/w·l/w) and provenance columns never enter metadata.input or the Given inputs block; mode labels appear only as task options. Raw MeltpoolNet process values are not exposed as naked numbers: velocity is normalized to mm/s, hatch/layer/beam dimensions to um, zero layer thickness is treated as unknown, and material composition/thermal properties are canonicalized from Akbari et al. Appendix A.9/A.10. This is a mode-only release based on HF revision 083cfafe758e72ecbfbaa049941668440300d516: it keeps the source-native mode task and removes the former maintainer-defined quantile-label task because those labels were not physical or literature-backed gold. T-AM2 emits 1,236 unique (query, annot) rows from 1,249 labeled source rows after folding exact duplicates with metadata.merged_paper_ids / provenance preserved. Hatch/layer completeness is metadata, not a classification filter. T-AM2 remains 5-way and imbalanced (spatter formation has only 8 rows). VED is retained only as a reference metric in metadata. The causal metadata scaffold is metadata.causal_features: normalized enthalpy and laser intensity for mode. metadata.causal_note states that VED is a powder-bed density metric, not the single-track melt-pool mode driver. metadata.cot_eligible is true iff normalized enthalpy is computable from P,V,beam,rho,Cp,k,Tm (1,084 true / 152 false). Eligible rows also carry metadata.cot_boundary_risk, computed by same-material k=5 nearest neighbors in standardized (log normalized enthalpy, log intensity) space with majority-vote mode prediction and tied votes marked risky; counts are 387 true / 697 false / 152 null. No reasoning/CoT is generated here. Metadata is serialized as strict JSON with NaN/Infinity rejected. No split file is attached; if splitting is needed, use source provenance carefully and do not randomly split rows. T-AM2 = AM task extension (metadata.taxonomy_note).

Provenance & licensing

Underlying source: MeltpoolNet — a meta-aggregation of ~hundreds of AM papers. Upstream license: no LICENSE (repo default = all-rights-reserved); meta-aggregation -> per-source clearance needed. Converted read-only into the unified schema; conversion + publish scripts live in AI4Manufacturing/forge_model (AM13_MeltpoolNet/ + publish/push_am_to_hf.py). Access is gated (manual approval); clear the upstream licence before any onward redistribution.