| --- |
| tags: |
| - smart-manufacturing |
| - sft |
| - industrial |
| - vision |
| license: other |
| pretty_name: "195" |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| extra_gated_prompt: >- |
| This dataset is released for **research use**. Access is reviewed and granted |
| **manually** by the maintainers. Please state your name, affiliation, and intended use. |
| --- |
| |
| # 195 |
|
|
| Cross-industry multimodal multiple-choice VQA benchmark (21 sectors; eval-only). Category **B**, task **T-B3**, in the unified Smart-Manufacturing SFT schema. |
|
|
| > The repository name is an internal task code. See **Provenance** below for the underlying dataset. |
|
|
| ## Records |
|
|
| **1,050** records (test=1050). |
|
|
| ## Unified SFT schema |
|
|
| | field | type | meaning | |
| |---|---|---| |
| | `query` | str | the question / instruction (model input) | |
| | `image` | Image | the input image (bytes embedded) | |
| | `annot` | str | the answer — for this dataset: the bare gold option letter (A-E; E is a 'cannot recognize' reject option), exactly the benchmark's answer. The 5 options are given in the query and in `metadata.choices`; the option text, the industrial `category`, and the parallel **Chinese** question/choices/answer are in `metadata` — see **Task & split** below | |
| | `reasoning` | null | no native CoT in these datasets | |
| | `cate` | "B" | SFT category | |
| | `task` | "T-xx" | unified task id | |
| | `metadata` | str (JSON) | split, provenance, `image_path`, `image_sha256` (dedup key) | |
| | `mask` | Image \| null | *(T-B1/T-B2 only)* the pixel ground-truth mask, bytes embedded | |
| | `masks` | list[Image] | *(D21 only)* multi-region masks | |
|
|
| ## Task & split |
|
|
| **What this is.** MME-Industry (arXiv 2501.16688) — a cross-industry multimodal **evaluation benchmark** for |
| MLLMs: 1,050 expert-authored multiple-choice questions over **21 industrial sectors** (50 each, e.g. power, |
| steel, chemical, textile, medical). Questions are deliberately **non-OCR** (they require domain reasoning, not |
| text reading). It is an eval-only benchmark scored by multiple-choice accuracy. |
|
|
| **Query & answer (this repo's SFT task).** `query` = the question followed by the 5 options `(A)…(E)` and the |
| instruction to answer with a single letter; `annot` = the **bare gold option letter** (`A`–`E`). Option `E` is a |
| `No corresponding features in the picture` reject option. The option text, the industrial `category`, and the |
| full set of options are in `metadata`. |
|
|
| **Bilingual.** The benchmark ships parallel **English and Chinese** versions (index-aligned). English is |
| authoritative for `query`/`annot`; the Chinese question / choices / category / answer are preserved in |
| `metadata` (`question_zh`, `choice_zh`, `category_zh`, `answer_zh`). The two languages agree on the answer for |
| 1,049 / 1,050 items; where they differ, the English answer is used. |
|
|
| **No mask.** This is VQA, not localization — there is no `mask` column. |
|
|
| **Split.** Single `test` split (1,050) — it is an evaluation benchmark, not a training set. |
|
|
| ## Provenance |
|
|
| Underlying dataset: **MME-Industry**. Upstream license: **Apache-2.0** (this card is `license: other`; respect the upstream terms). Converted read-only from the raw source into the unified schema; conversion script: `195/convert_d95.py`, published with `publish/push_to_hf.py`, both in [`AI4Manufacturing/forge_model`](https://github.com/AI4Manufacturing/forge_model). |
|
|
| ## Overlap / de-duplication (§8) |
|
|
| Evaluation benchmark (single `test` split); expert-authored, non-OCR — no image overlap with the other datasets. Each record carries `metadata.image_sha256` so overlapping images can be kept entirely on one side of a train/eval split. |
|
|
