taskweft-fbd-editscore-train

Intents paired with IEC 61131-3 Function Block Diagrams that carry them out, as an EditScore-shaped corpus.

What it is for

Each intent has a root row, three ranked candidate diagrams and a score row per candidate. Rank1 is the reference diagram. Rank3 compiles and does the wrong thing. Rank5 is a diagram the compiler refuses. The scores come from the compiler and a runner that performed the plan. Every row is built from a template and a seed, so the labels are true by construction and the corpus regenerates from the seeds. The train split trains. The test split holds out every tenth seed of the same distribution and gates the trained model. The evaluation split holds out the write_then_count family, and is never trained or tuned on.

The editscore_root config holds the intent, template, seed, frame, the block kinds and signatures the reference uses, and the input traces. The editscore_candidates config holds each candidate's compiler text form, PLCopen XML, plan and runner result, with a sha256 of its source text. The editscore_scores config holds parse, compile, run and effect-match flags, step and wall-time measures, and the refusal. The default editscore config joins the three. Every artifact is in the parquet. An artifact a family does not produce is an empty string, never a null. A manifest.json sits at the repo root.

Provenance

• Generator: v-sekai-fabric/taskweft-fbd-teacher.
• Compiler: taskweft-fbd-compiler at commit d3279ec.

Use

from datasets import load_dataset
ds = load_dataset("chibifire/taskweft-fbd-editscore-train", "editscore", split="train")

Or fetch the files with hf download chibifire/taskweft-fbd-editscore-train --repo-type dataset.

Licence

MIT. See the MIT License.