| """Concept token sets C_t for TOFU forget10 authors (proposal §1.2). |
| |
| For each author: tokenization variants of name parts (leading-space / |
| no-space, first subword token of each). Prompt partition: name_absent |
| questions (author name not in question text) are the H1 primary signal; |
| name_present questions are reported separately (copy confound). |
| """ |
|
|
| import json |
| import re |
|
|
| FORGET10_AUTHORS = [ |
| "Hsiao Yun-Hwa", "Carmen Montenegro", "Elvin Mammadov", "Rajeev Majumdar", |
| "Jad Ambrose Al-Shamary", "Adib Jarrah", "Ji-Yeon Park", "Behrouz Rohani", |
| "Wei-Jun Chen", "Tae-ho Park", "Hina Ameen", "Xin Lee Williams", |
| "Moshe Ben-David", "Kalkidan Abera", "Takashi Nakamura", "Raven Marais", |
| "Aysha Al-Hashim", "Edward Patrick Sullivan", "Basil Mahfouz Al-Kuwaiti", |
| "Nikolai Abilov", |
| ] |
|
|
|
|
| def name_parts(full_name: str) -> list[str]: |
| """'Hsiao Yun-Hwa' -> ['Hsiao', 'Yun-Hwa', 'Yun', 'Hwa'] (space and hyphen splits).""" |
| parts = full_name.split() |
| for p in list(parts): |
| if "-" in p: |
| parts.extend(seg for seg in p.split("-") if len(seg) > 2) |
| return parts |
|
|
|
|
| def concept_token_ids(tokenizer, full_name: str) -> list[int]: |
| """First subword token of each name variant, leading-space and bare.""" |
| ids = [] |
| variants = [full_name] + name_parts(full_name) |
| for v in variants: |
| for s in (f" {v}", v): |
| t = tokenizer(s, add_special_tokens=False).input_ids |
| if t: |
| ids.append(t[0]) |
| return sorted(set(ids)) |
|
|
|
|
| def author_in_text(full_name: str, text: str) -> bool: |
| """Any name part (len>2) appears in text (case-insensitive, word-ish match).""" |
| low = text.lower() |
| for p in name_parts(full_name): |
| if len(p) > 2 and re.search(re.escape(p.lower()), low): |
| return True |
| return False |
|
|
|
|
| def load_forget10_rows(): |
| """Rows of forget10 with author index/name and name_present flag.""" |
| from datasets import load_dataset |
|
|
| ds = load_dataset("locuslab/TOFU", "forget10")["train"] |
| rows = [] |
| for i in range(len(ds)): |
| author = FORGET10_AUTHORS[i // 20] |
| rows.append( |
| { |
| "idx": i, |
| "author_idx": i // 20, |
| "author": author, |
| "question": ds[i]["question"], |
| "answer": ds[i]["answer"], |
| "name_present": author_in_text(author, ds[i]["question"]), |
| } |
| ) |
| return rows |
|
|
|
|
| if __name__ == "__main__": |
| import transformers |
|
|
| tok = transformers.AutoTokenizer.from_pretrained("/workspace/models/tofu_1b_full") |
| rows = load_forget10_rows() |
| n_absent = sum(not r["name_present"] for r in rows) |
| print(f"forget10: {len(rows)} rows, {n_absent} name-absent") |
| from collections import Counter |
|
|
| per_author = Counter(r["author"] for r in rows if not r["name_present"]) |
| for a in FORGET10_AUTHORS: |
| ids = concept_token_ids(tok, a) |
| toks = [tok.decode([t]) for t in ids] |
| print(f"{a:26s} absent={per_author.get(a, 0):2d} C_t={toks}") |
| json.dump(rows, open("/workspace/jspace-unlearning/results/forget10_rows.json", "w"), indent=1) |
|
|
