File size: 3,632 Bytes
a3773e2 | 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 | #!/usr/bin/env python3
"""Descriptive source margins, perturbations, and paired rehearsal effects."""
from __future__ import annotations
import csv
import itertools
import numpy as np
from llm_memory_editability.hebbian_future import ART, CONFIG, RESULTS, read, write
def main():
cfg = read(CONFIG)["chains"]
rows, world_effects, groups = [], [], []
for directory in sorted((RESULTS / "chains").glob("w*-r*-d*")):
if not (directory / "complete.json").exists():
continue
meta = read(directory / "complete.json")
key = {k: meta[k] for k in ("world", "rho", "depth")}
parent = read(directory / "learning.json")[-1]
stages = [("parent", parent)]
stages += [
(arm["arm"], arm["trajectory"][-1]) for arm in read(directory / "rehearsals.json")
]
for arm, row in stages:
record = {**key, "arm": arm, **row["summary"]}
for kind in ("member", "root", "home", "composite"):
margins = np.asarray(row["rows"][kind]["margin"])
record[kind + "_margin_median"] = float(np.median(margins))
record[kind + "_margin_q10"] = float(np.quantile(margins, 0.1))
rows.append(record)
for arm in cfg["rehearsal_arms"]:
final = next(
r for r in rows if all(r[k] == v for k, v in key.items()) and r["arm"] == arm
)
for metric in (
"held_composite",
"held_common_conflict_composite",
"held_two_step",
"member_accuracy",
"root_accuracy",
"home_accuracy",
):
world_effects.append(
{
**key,
"arm": arm,
"metric": metric,
"change_from_parent": final[metric] - parent["summary"][metric],
}
)
metrics = [
"held_composite",
"held_common_conflict_composite",
"held_two_step",
"member_accuracy",
"root_accuracy",
"home_accuracy",
"member_margin_median",
"root_margin_median",
"home_margin_median",
]
for depth, rho, arm in itertools.product(
cfg["layers"], cfg["correlations"], ["parent", *cfg["rehearsal_arms"]]
):
part = [r for r in rows if r["depth"] == depth and r["rho"] == rho and r["arm"] == arm]
if part:
groups.append(
{
"depth": depth,
"rho": rho,
"arm": arm,
"worlds": len(part),
**{k: float(np.mean([r[k] for r in part])) for k in metrics},
}
)
for name, values in [
("rehearsal-detail.csv", rows),
("rehearsal-effects.csv", world_effects),
("rehearsal-aggregates.csv", groups),
]:
with (ART / name).open("w") as handle:
writer = csv.DictWriter(handle, fieldnames=list(values[0]))
writer.writeheader()
writer.writerows(values)
write(
ART / "rehearsal-summary.json",
{
"aggregates": groups,
"effects": world_effects,
"interpretation": (
"Raw margins are descriptive within model/arm, not scale invariant robustness. "
"Use perturbation curves separately. Rehearsal changes parameters broadly, "
"not margin alone."
),
},
)
if __name__ == "__main__":
main()
|