File size: 4,235 Bytes
2eb3475 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 | """Cross-interaction features.
Each module exposes a plain function ``extract_<name>(model, input_ids,
attention_mask, cache, pred_label=None) -> (features, columns)``. The
extractor entry-point dispatches by name.
PHASE_OF[name] tells the orchestrator whether a feature needs gradients
enabled on the forward pass. Phase 3 and 4 features set this to True;
Phase 0/1/2 use the cheap no-grad ``ForwardCache``.
"""
from .lgap import extract_lgap
from .scibd_1p import extract_scibd_1p
from .rscd import extract_rscd
from .owps import extract_owps
from .kcs import extract_kcs
from .tcvcg import extract_tcvcg
from .tcvcg_rand import extract_tcvcg_rand
from .scibd_2p import extract_scibd_2p
from .gclp import extract_gclp
from .jcsi import extract_jcsi
from .prep import extract_prep
from .eapt import extract_eapt
# Sprint 1: PD-on-attention topology features (clean cross-attention)
from .pd_hdg import extract_pd_hdg
from .pbd import extract_pbd
from .rpd import extract_rpd
from .zap import extract_zap
from .rcp import extract_rcp
from .hzap import extract_hzap
from .rzap import extract_rzap
from .ttp import extract_ttp
from .aaa import extract_aaa
from .mhp import extract_mhp
from .fzp import extract_fzp
from .cler import extract_cler
from .idt import extract_idt
from .rsc import extract_rsc
from .art import extract_art
from .clsph import extract_clsph
from .ctcka import extract_ctcka
from .ahj import extract_ahj
from .mwp import extract_mwp
from .vas import extract_vas
from .btl import extract_btl
from .laed import extract_laed
from .lid import extract_lid
from .tbz import extract_tbz
EXTRACTORS = {
# Phase 0 — hook-only
"lgap": extract_lgap,
"scibd_1p": extract_scibd_1p,
# Phase 1 — 1x forward + light post-processing
"rscd": extract_rscd,
"owps": extract_owps,
"kcs": extract_kcs,
# Phase 2 — setup-heavy, cheap inference
"tcvcg": extract_tcvcg,
"tcvcg_rand": extract_tcvcg_rand,
"scibd_2p": extract_scibd_2p,
# Phase 3 — gradient-based
"gclp": extract_gclp,
"jcsi": extract_jcsi,
# Phase 4 — expensive
"prep": extract_prep,
"eapt": extract_eapt,
# Sprint 1 — PD-on-attention (clean topology)
"pd_hdg": extract_pd_hdg,
"pbd": extract_pbd,
"rpd": extract_rpd,
"zap": extract_zap,
"rcp": extract_rcp,
# Sprint 1.5 — head-axis zigzag (this paper, admissible cross-feature hunt)
"hzap": extract_hzap,
"rzap": extract_rzap,
# Token-trajectory persistence (hidden-state substrate)
"ttp": extract_ttp,
"aaa": extract_aaa,
"mhp": extract_mhp,
"fzp": extract_fzp,
# Sprint-2 — Tier-A admissible cross-features (this paper).
"cler": extract_cler,
"idt": extract_idt,
"rsc": extract_rsc,
"art": extract_art,
"clsph": extract_clsph,
# Sprint-3 — cross-substrate (user directive: combine activations + attention)
"ctcka": extract_ctcka,
"ahj": extract_ahj,
# Sprint-4 — orthogonal substrate cross-features
"mwp": extract_mwp,
"vas": extract_vas,
"btl": extract_btl,
"laed": extract_laed,
"lid": extract_lid,
"tbz": extract_tbz,
}
PHASE_OF = {
"lgap": 0, "scibd_1p": 0,
"rscd": 1, "owps": 1, "kcs": 1,
"tcvcg": 2, "tcvcg_rand": 2, "scibd_2p": 2,
"gclp": 3, "jcsi": 3,
"prep": 4, "eapt": 4,
# Sprint 1 features form a separate phase 5 ("PD-on-attention").
"pd_hdg": 5, "pbd": 5, "rpd": 5, "zap": 5, "rcp": 5,
"hzap": 6,
"rzap": 6,
"ttp": 6,
"aaa": 6,
"mhp": 6,
"fzp": 7,
"cler": 7, "idt": 7, "rsc": 7, "art": 7, "clsph": 7,
"ctcka": 8, "ahj": 8,
"mwp": 9, "vas": 9,
"btl": 10,
"laed": 10,
"lid": 10,
}
NEEDS_GRADS = {"gclp", "jcsi", "eapt"} # features that compute their own bwd
NEEDS_NO_CACHE = {"prep"} # features that do their own forwards
|