File size: 4,730 Bytes
ead9072 | 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 | """Unified per-site table for the 8 focus names (+ bare_Alice control).
Columns: name, variant, pos, update, doc_idx, block_first_mention(same-id),
standalone, prev_cat, next_piece, ctx_multiplicity, story_reg, dialog,
she_in_prefix, girl_in_prefix. Written as parquet if pyarrow available, else csv.gz.
"""
import json, os, sys
from collections import Counter
import numpy as np
REPO = "/mnt/nlpgpu-io1/data/maggie/explain-m0de-collapse/replicate"
sys.path.insert(0, REPO)
BASE = os.path.join(REPO, "out/alice_vs_megan")
K = 16; BLOCK = 1024; W = 131072; NL = 198; EOS = 100257
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(
os.path.join(REPO, "data/olmo3_prefix_hfstream_50b/tokenizer"), trust_remote_code=True)
sites = np.load(os.path.join(BASE, "sites/sites_all.npz"))
ctxs = np.load(os.path.join(BASE, "sites/contexts_all.npz"))
flags = np.load(os.path.join(BASE, "sites/block_register_flags.npz"))
eos = np.load(os.path.join(BASE, "sites/eos_positions.npy"))
h0 = json.load(open(os.path.join(BASE, "sites/h0_counts.json")))
dec_cache = {}
def dec(i):
i = int(i)
if i not in dec_cache:
dec_cache[i] = tok.decode([i])
return dec_cache[i]
HONOR = {"mrs","mr","ms","miss","aunt","auntie","aunty","lady","princess","queen",
"sister","grandma","grandmother","mother","mom","mama","cousin","dr",
"madam","madame","saint","st","mrs.","ms.","dr.","st.","dame"}
PREPS = {"about","with","to","from","for","of","as","at","on","in","like"}
QUOTES = set('"“”‘’\'`«»')
def prev_cat(pid):
s = dec(pid)
if pid == EOS: return "doc_start"
if pid == NL or s.endswith("\n"): return "newline"
t = s.strip(); tl = t.lower()
if not t: return "space"
if tl in ("named","called"): return "named_called"
if tl == "dear": return "dear"
if tl == "and": return "and"
if tl in HONOR: return "honorific_kin"
if t[-1] in ".!?…": return "sentence_end"
if t[-1] in QUOTES: return "quote_adj"
if t[-1] == ":": return "colon"
if t[-1] == ",": return "comma"
if tl == "by": return "by"
if tl in PREPS: return "prep"
if t[0].isdigit(): return "digit"
if s[0] == " " and t[0].isupper(): return "cap_word"
if s[0] == " " and t[0].islower(): return "lower_word"
if t[0].islower(): return "glue_lower"
return "other"
def is_standalone(nid):
s = dec(nid); t = s.strip()
if not s: return True
if s[0].isalnum(): return False # glued piece
cap = bool(t) and s[0] == " " and t[0].isupper() and t.replace("-", "").replace(".", "").isalpha()
return not cap
recs = []
FOCUS = ["Alice", "Mary", "Sarah", "Maria", "Julia", "Megan", "Karen", "Ann"]
for nm in FOCUS + ["bare_Alice"]:
lab = nm if nm.startswith("bare") else f"sp_{nm}"
p = sites[f"pos_{lab}"]; C = ctxs[f"ctx_{lab}"]
if not len(p): continue
seqlen = len(h0["patterns"][lab]["ids"])
cs = ((p - 1) // BLOCK) * BLOCK
prev_same = np.concatenate(([np.int64(-1)], p[:-1]))
fm = prev_same < cs
rows = [tuple(r) for r in C.tolist()]
mult = Counter(rows)
doc_idx = np.searchsorted(eos, p)
has_flags = (not nm.startswith("bare")) and (f"{nm}_story_reg" in flags.files)
for i in range(len(p)):
recs.append({
"name": nm.replace("bare_", ""), "variant": "bare" if nm.startswith("bare") else "sp",
"pos": int(p[i]), "update": int(p[i] // W), "doc_idx": int(doc_idx[i]),
"first_mention_block": bool(fm[i]),
"standalone": is_standalone(C[i, K + seqlen]),
"prev_cat": prev_cat(int(C[i, K - 1])),
"next_piece": dec(C[i, K + seqlen])[:20],
"ctx_multiplicity": int(mult[rows[i]]),
"story_reg": bool(flags[f"{nm}_story_reg"][i]) if has_flags else None,
"dialog_prefix": bool(flags[f"{nm}_dialog"][i]) if has_flags else None,
"she_in_prefix": bool(flags[f"{nm}_she"][i]) if has_flags else None,
"girl_in_prefix": bool(flags[f"{nm}_girl"][i]) if has_flags else None,
})
out_base = os.path.join(BASE, "per_site_table")
try:
import pandas as pd
df = pd.DataFrame(recs)
try:
df.to_parquet(out_base + ".parquet", index=False)
print("wrote", out_base + ".parquet", len(df), "rows")
except Exception as e:
df.to_csv(out_base + ".csv.gz", index=False, compression="gzip")
print("parquet failed (", e, ") -> wrote csv.gz", len(df))
except ImportError:
import csv, gzip
with gzip.open(out_base + ".csv.gz", "wt", newline="") as f:
w = csv.DictWriter(f, fieldnames=list(recs[0]))
w.writeheader(); w.writerows(recs)
print("wrote", out_base + ".csv.gz", len(recs), "rows")
|