Data points look like X_i,y_i = (Gamma |- C, rule) #SAMPLES = (20k training, 2k test) All trees have a maximum number of 100 nodes, represented as a 100 x 2 matrix

*_ctxs.npz

The trees in Gamma. Dim = #SAMPLES x 5 x 100 x 2

*_goals.npz

The tree C. Dim = #SAMPLES x 100 x 2

*_n.npz

The labels of the nodes in Gamma (resp. C). Dim = #SAMPLES x 5 x 100 x 6 (resp. #SAMPLES x 100 x 6). Last dim layed out as (VAR, TYVAR, TYAPP, CONST, COMB, ABS) in [128] x [128] x [297] x [297] x [1] x [1], since there are 128 ascii characters and 297 hol-light constants.

*_meta.npz

Data about the actual sizes input. Dim = #SAMPLES x 2. Last dim layouted as (|Gamma|, |C|, proof_index). proof_index is the index of the theorem.

TODO: NEED TO INCLUDE sizes of each A in Gamma: |A|.

*_labels.npz

One hot embedding of the rules. Dim = #SAMPLES x 13. Indices:

Prefl(tm) -> 0 Ptrans(p1,p2) -> 1 Pmkcomb(p1,p2) -> 2 Pabs(p1,tm) -> 3 Pbeta(tm) -> 4 Passume(tm) -> 5 Peqmp(p1,p2) -> 6 Pdeduct(p1,p2) -> 7 Pinst(p1,insts) -> 8 Pinstt(p1,insts) -> 9 Paxiom(tm) -> 10 Pdef(tm,name,ty) -> 11 Pdeft(p1,tm,name,ty) -> 12