llm-scfg

Synthetic parallel-language data generated from paired synchronous context-free grammars (SCFGs). Each experiment config holds three splits:

• grammars — one row per grammar, with the SCFG source, lexicons, and typological knobs for both language a and language b.
• samples — generated (left, right) sentence pairs joined to a grammar by grammar_name.
• shots — held-out example pools for few-shot prompting. This split is empty for experiments that do not define shots_*.jsonl files.

Loading

from datasets import load_dataset

grammars = load_dataset("<owner>/llm-scfg", "agreement_exp", split="grammars")
samples  = load_dataset("<owner>/llm-scfg", "agreement_exp", split="samples")
shots    = load_dataset("<owner>/llm-scfg", "fewshot_exp", split="shots")

Join samples.grammar_name == grammars.name for grammar-conditioned analysis.

Configs

• agreement_exp
• complexity_exp
• fewshot_exp
• orthography_exp
• size_exp
• wordorder_exp

Schemas

grammars

Stable scalar knobs are lifted to a_* / b_* columns; lexicons stay as list columns. The full original grammar JSON (including paradigms, agreement suffix tables, and any experiment-specific extras) is preserved verbatim in raw_json.

• experiment (string): matches the config name.
• name (string): grammar hash, used as the join key.
• grammar_str (string): SCFG source.
• n_rules, n_words (int64): grammar stats.
• has_agreement_metadata (bool): true on configs that record agreement_metadata.
• agreement_metadata_json (string): full agreement metadata object where present.
• lexical_frequency_metadata_json (string): lexical-frequency settings for the paired grammar.
• a_*, b_* scalars (typed): head_initial, spec_initial, syllable_structure, rng_seed, lexical_frequency_profile, lexical_frequency_exponent, lexical_frequency_length_unit, etc. — see column names in the parquet schema.
• a_*, b_* lists (list<string>): verbs, nouns, propns, prons, adjs, det_def, det_indef, comps, tenses, asps, agreement_axes, gender_values.
• raw_json (string): full grammar JSON for round-tripping.

Fields not present on a given experiment (e.g. a_orthography on complexity_exp) are null.

samples

• experiment (string): matches the config name.
• source_file (string): original JSONL basename.
• row_index (int64): zero-based line index within source_file.
• grammar_name (string): join key into grammars.
• left, right (string): parallel surface strings.
• left_phonetic, right_phonetic (string): phonetic forms.
• left_tree, right_tree (string): parse trees.
• depth, min_depth, max_depth (int64): derivation-depth metadata.
• rng_seed (int64): per-sample seed.
• subject_person, subject_number, verb_person, verb_number (string): agreement features; null where not applicable.
• agreement_ok (bool): null where not applicable.
• agreement_trace (string): null where not applicable.
• possible_right, possible_right_phonetic (list<string>): alternates considered during generation; null where not recorded.
• raw_json (string): full original sample JSON for round-tripping.

shots

Same schema as samples, but rows come from shots_*.jsonl files. These rows are held-out few-shot example pools; they are not evaluation samples.

Notes

• Experiments tagged with has_agreement_metadata=true were generated with the agreement-aware pipeline; samples on those configs populate the subject_* / verb_* / agreement_* columns.
• Some grammars in agreement_exp have no shipped samples file (the original artifacts were multi-GB and excluded from the public release). Those rows still appear in the grammars split; see manifest.json for the list.