anand-bala/automata-embeddings

Craftax LTLf Brzozowski Dataset (bz-3.0)

A Brzozowski-derivative dataset built from 67 Craftax achievement tasks expressed as LTLf formulas. Each task is compiled to a minimal DFA via morphata and unrolled into per-residual (Brzozowski quotient) records, DFA transitions, sampled episodes, and non-equivalence pairs.

At a glance

• Tasks: 67 Craftax achievement formulas (LTLf over per-tile predicates).
• Splits: train (47), val (10), test (10). One large outlier (enchant_armour, 3,186 DFA states) sits in train; the largest task (enchant_sword, 16,888 states) is held out in test to probe size-extrapolation of learned embeddings.
• Episode sampling configs: two sampling tags, both with prefix length 64 and episode length 16:
  • dataset-v3-prefix64-walk16-len16-seed0 (random walk depth 16)
  • dataset-v3-prefix64-walk32-len16-seed0 (random walk depth 32)

Directory layout

.
|-- craftax_tasks_{train,val,test}.parquet
|-- craftax_residuals_{train,val,test}.parquet
|-- craftax_transitions_{train,val,test}.parquet
|-- craftax_<sampling_tag>_episodes_{train,val,test}.parquet
|-- craftax_<sampling_tag>_neq_{train,val,test}.parquet
|-- splits.json                       # {train: [task_id], val: [...], test: [...]}
`-- tasks/<task_id>/                  # per-task morphata sidecars
    |-- alphabet.json
    |-- formula.txt
    |-- min_dfa.bdd.json              # canonical min-DFA (BDD-encoded)
    |-- min_dfa.bdd.json.meta.json
    |-- composition_map.json          # conjunctive tasks only
    `-- factor_<i>.bdd.json[.meta.json]

Every parquet file is keyed by task_id (string, prefixed craftax_<name>), which joins back to tasks/<task_id>/ and across all other tables.

Table schemas

tasks (1 row per task)

Column	Type	Description
task_id	string	Unique task identifier (craftax_<name>).
name	string	Bare task name (e.g. enchant_sword).
formula_str	string	LTLf source formula.
alphabet_kind	string	Alphabet encoding family (e.g. one_hot).
predicates	list<string>	Atomic propositions used by the formula.
symbols	list<list<str>>	Alphabet symbols, each a subset of predicates.
num_states	int64	Number of states in the minimal DFA.

residuals (1 row per (task_id, residual_id))

Column	Type	Description
task_id	string	Foreign key into tasks.
residual_id	int64	DFA state index (the Brzozowski residual identifier).
boolexpr	string	Symbolic label of the residual (canonical state expression).
accepting	bool	true iff this residual is an accepting DFA state.
num_states	int64	DFA state count of the parent task (repeated for convenience).

transitions (1 row per DFA edge)

Column	Type	Description
task_id	string	Foreign key into tasks.
src	int64	Source residual_id.
symbol	int64	Index into the task's symbols list.
dst	int64	Destination residual_id.

episodes_<sampling_tag> (1 row per sampled trajectory)

Column	Type	Description
task_id	string	Foreign key into tasks.
episode_idx	int64	Per-task episode index.
policy	string	Sampling policy (e.g. prefix_then_random, random_from_residual).
symbol_seq	list<int64>	Symbol indices visited (len == episode_length).
residual_seq	list<int64>	residual_id sequence, aligned to symbol_seq.
accepting_seq	list<bool>	Per-step accepting flag of the visited residual.

neq_<sampling_tag> (1 row per non-equivalence witness)

Column	Type	Description
task_id	string	Foreign key into tasks.
residual_a	int64	First residual_id of the distinguished pair.
residual_b	int64	Second residual_id of the distinguished pair.
distinguishing_word	list<int64>	Symbol-index sequence accepted from one residual and rejected from the other.

splits.json

Flat JSON mapping each split name to its list of task_ids:

{
  "train": ["craftax_cast_iceball", "..."],
  "val":   ["craftax_...", "..."],
  "test":  ["craftax_enchant_sword", "..."]
}

Per-task sidecars (tasks/<task_id>/)

The tasks/ subtree carries the canonical morphata representation of each DFA, written by brzozowski_dataset and consumed by downstream tooling (e.g. RAD comparison sweeps). The files are not partitioned across splits; they sit alongside the per-split parquet shards and are looked up by task_id.

File	Description
alphabet.json	Predicate names and the enumerated symbol set.
formula.txt	Source LTLf formula.
min_dfa.bdd.json	Canonical minimal DFA (BDD-encoded transition function).
min_dfa.bdd.json.meta.json	Metadata (variable order, accepting set, etc.).
composition_map.json	Conjunctive tasks only: factor-to-product state composition map.
factor_<i>.bdd.json[.meta]	Conjunctive tasks only: per-factor BDD DFA components.