File size: 7,968 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | #!/usr/bin/env python3
"""Explicit execution entry point for the prospective Hebbian study."""
import argparse
import importlib.metadata
import json
import platform
import re
import subprocess
import sys
from llm_memory_editability.hebbian_learning import (
ARTIFACTS,
CONFIG_PATH,
DATA,
calibrate,
config,
now,
read_json,
sha256,
write_json,
)
def prepare_sources():
import requests
from transformers import AutoTokenizer
from llm_memory_editability.hebbian_data import clean_candidates
cfg = config()
source = DATA / "source"
source.mkdir(parents=True, exist_ok=True)
def fetch(url, path):
if path.exists():
return
path.parent.mkdir(parents=True, exist_ok=True)
with requests.get(url, stream=True, timeout=(30, 180)) as response:
response.raise_for_status()
tmp = path.with_name(path.name + ".part")
with tmp.open("wb") as f:
for block in response.iter_content(4 * 1024 * 1024):
f.write(block)
tmp.replace(path)
fetch("https://rome.baulab.info/data/dsets/counterfact.json", source / "counterfact.json")
model = cfg["model"]
fetch(
f"https://huggingface.co/api/models/{model['id']}/revision/{model['revision']}",
source / "model-info.json",
)
model_info = read_json(source / "model-info.json")
assert model_info["sha"] == model["revision"]
for item in model_info["siblings"]:
name = item["rfilename"]
if name != ".gitattributes":
fetch(
f"https://huggingface.co/{model['id']}/resolve/{model['revision']}/{name}",
source / "qwen3-0.6b-base" / name,
)
rev = cfg["author_revision"]
for name in [
"src/hebbian/methods/hebbian/model.py",
"src/hebbian/methods/hebbian/construction.py",
"LICENSE",
"README.md",
]:
fetch(
f"https://raw.githubusercontent.com/HazyResearch/hebbian-mlps/{rev}/{name}",
source / "hebbian-mlps" / name,
)
fetch("https://huggingface.co/api/datasets/Salesforce/wikitext", source / "wikitext-info.json")
wiki_info = read_json(source / "wikitext-info.json")
tokenizer = AutoTokenizer.from_pretrained(source / "qwen3-0.6b-base", local_files_only=True)
import pyarrow.parquet as pq
texts = {}
for split in ["train", "test"]:
name = f"wikitext-2-raw-v1/{split}-00000-of-00001.parquet"
path = source / name
fetch(
f"https://huggingface.co/datasets/Salesforce/wikitext/resolve/{wiki_info['sha']}/{name}",
path,
)
lines = pq.read_table(path).column("text").to_pylist()
documents, current = [], []
for line in lines:
if re.fullmatch(r"\s*= [^=]+ =\s*", line) and current:
documents.append("".join(current))
current = []
current.append(line)
if current:
documents.append("".join(current))
segments = []
for doc_id, doc in enumerate(documents):
tokens = tokenizer(doc, add_special_tokens=False)["input_ids"]
for offset in range(0, len(tokens) - 127, 128):
ids = tokens[offset : offset + 128]
segments.append(
{
"document_id": doc_id,
"token_offset": offset,
"document_sha256": __import__("hashlib").sha256(doc.encode()).hexdigest(),
"input_ids": ids,
"loss_mask": [0] + [1] * 127,
"answer_start": 1,
"source_split": split,
}
)
if len(segments) == 128:
break
if len(segments) == 128:
break
assert len(segments) == 128
texts[split] = segments
write_json(DATA / "text.json", texts)
files = {
str(p.relative_to(source)): sha256(p)
for p in source.rglob("*")
if p.is_file() and not p.name.endswith(".part")
}
write_json(
ARTIFACTS / "source-manifest.json",
{
"time": now(),
"model_revision": model["revision"],
"author_revision": rev,
"wikitext_revision": wiki_info["sha"],
"sha256": files,
"model_license": "Apache-2.0",
"author_license": (source / "hebbian-mlps/LICENSE").read_text(),
"wikitext_card": wiki_info.get("cardData"),
"config_sha256": sha256(CONFIG_PATH),
},
)
if not (DATA / "candidates.json").exists():
clean_candidates(tokenizer)
packages = {}
for name in [
"torch",
"transformers",
"tokenizers",
"datasets",
"huggingface-hub",
"numpy",
"matplotlib",
"safetensors",
"pyarrow",
]:
try:
packages[name] = importlib.metadata.version(name)
except importlib.metadata.PackageNotFoundError:
packages[name] = "not installed"
write_json(
ARTIFACTS / "environment-lock.json",
{
"time": now(),
"python": sys.version,
"executable": sys.executable,
"platform": platform.platform(),
"packages": packages,
"dtype": "float32",
"attention": "eager",
"tf32": False,
"deterministic_algorithms": True,
"CUBLAS_WORKSPACE_CONFIG": ":4096:8",
"gpu_query": subprocess.check_output(
[
"nvidia-smi",
"--query-gpu=index,name,uuid,ecc.errors.uncorrected.volatile.total",
"--format=csv",
],
text=True,
),
"excluded_gpu_indices": [0],
},
)
(ARTIFACTS / "environment-freeze.txt").write_text(
subprocess.check_output([sys.executable, "-m", "pip", "freeze"], text=True)
)
def main():
parser = argparse.ArgumentParser()
parser.add_argument(
"command",
choices=[
"prepare",
"calibrate",
"baseline",
"audit-data",
"lock-data",
"freeze-b",
"preflight",
"diagnose",
"form",
"adapt",
],
)
parser.add_argument("--device", default="cuda:3")
parser.add_argument("--split", choices=["dev", "eval"], default="dev")
parser.add_argument("--run-id", default="hebbian-learning-v1")
parser.add_argument("--shard", type=int, default=0)
parser.add_argument("--shards", type=int, default=1)
args = parser.parse_args()
if args.command == "prepare":
prepare_sources()
elif args.command == "calibrate":
calibrate()
elif args.command == "baseline":
from llm_memory_editability.hebbian_model import run_baseline
run_baseline(args.device, args.shard, args.shards)
elif args.command == "audit-data":
from llm_memory_editability.hebbian_data import semantic_audit
semantic_audit()
elif args.command == "freeze-b":
from llm_memory_editability.hebbian_statistics import freeze_b
freeze_b()
elif args.command == "lock-data":
from llm_memory_editability.hebbian_data import lock_data
rows = []
for p in sorted(DATA.glob("baseline-shard-*.jsonl")):
rows.extend(json.loads(line) for line in p.read_text().splitlines())
assert len({r["case_id"] for r in rows}) == len(rows)
write_json(DATA / "baseline.json", sorted(rows, key=lambda r: r["case_id"]))
print(lock_data())
else:
from llm_memory_editability.hebbian_train import dispatch
dispatch(args)
if __name__ == "__main__":
main()
|