taskweft-fbd-editscore-train

Intents and the IEC 61131-3 Function Block Diagrams that carry them out, as an EditScore-shaped corpus: one root row per intent, three candidates per row (rank1 the reference diagram, rank3 one that compiles and does the wrong thing, rank5 one the compiler refuses), and one score row per candidate from the compiler and a runner that performed the plan. Every row is constructed from a template and a seed, so the labels are true by construction and the corpus regenerates from the seeds. Controls were asserted on every row before the emit.

Three splits. train trains. test is the same distribution with every tenth seed held out: the gate after training. evaluation is whole held-out families and block kinds (write_then_count), never trained or tuned on.

What a row carries

Every artefact is in the parquet; nothing points outside it.

fbd_root, one row per intent: key, intent, template_id, seed, frame_id, blocks (the block kinds and signatures the reference diagram uses), traces (the input traces as JSON strings, empty where the family has none), and the EditScore stub columns.

fbd_candidates, three per row: candidate and rank, fbd_text (the compiler's text form, verbatim), fbd_xml (the PLCopen XML, verbatim), plan_json and result_json (what the compiler planned and what the runner returned, where the family has a runner), fbd_sha (sha256 of the text the candidate was built from), fbd_path (where the writer kept it on disk) and provenance. An artefact a family does not produce is an empty string, never a null.

fbd_scores, three per row: parses, compiles, runs, effect_matches, steps, wall_ms, refusal.

fbd is the joined view and the default config: the root row with its candidates, each carrying its scores.

Rows: 5000 (3375 train, 375 test, 1250 evaluation). Templates: {"write": 1250, "write_then_read": 1250, "run_print": 1250, "write_then_count": 1250}. Compiler: taskweft-fbd-compiler d3279ec. Source: v-sekai-fabric/taskweft-fbd-teacher.