alice_vs_megan/code/analyze.py

"""H1-H4 contrastive analysis of name sites in updates [50000,55000).

Consumes out/alice_vs_megan/sites/* produced by scan_extract.py.
Produces out/alice_vs_megan/stats/*.json, examples_*.txt, tables_*.md.
"""
import json, os, sys, math, random
from collections import Counter, defaultdict
import numpy as np

REPO = "/mnt/nlpgpu-io1/data/maggie/explain-m0de-collapse/replicate"
sys.path.insert(0, REPO)
from scripts.name_lists import GIRLS

BASE = os.path.join(REPO, "out/alice_vs_megan")
SITES = os.path.join(BASE, "sites")
STATS = os.path.join(BASE, "stats")
os.makedirs(STATS, exist_ok=True)

W = 131072
U0, U1 = 50000, 55000
BLOCK = 1024
EOS = 100257
NL = 198
K = 16
FOCUS = ["Alice", "Maria", "Mary", "Julia", "Ann", "Megan", "Karen", "Sarah"]

from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained(
    os.path.join(REPO, "data/olmo3_prefix_hfstream_50b/tokenizer"), trust_remote_code=True)

_dec_cache = {}
def dec(i):
    i = int(i)
    if i not in _dec_cache:
        _dec_cache[i] = tok.decode([i])
    return _dec_cache[i]

sites = np.load(os.path.join(SITES, "sites_all.npz"))
ctxs = np.load(os.path.join(SITES, "contexts_all.npz"))
eos_pos = np.load(os.path.join(SITES, "eos_positions.npy"))
h0 = json.load(open(os.path.join(SITES, "h0_counts.json")))
labels = [k[4:] for k in sites.files]  # strip "pos_"

def pos_of(lab): return sites[f"pos_{lab}"]
def ctx_of(lab): return ctxs[f"ctx_{lab}"]

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"

SLOT_CUES = {"named_called","dear","and","honorific_kin","sentence_end","quote_adj",
             "colon","doc_start","newline"}

def next_cat(nid):
    s = dec(nid)
    if nid == EOS: return "doc_end"
    if nid == NL or s.startswith("\n"): return "newline"
    t = s.strip()
    if not t: return "space"
    if t in ("'s","’s"): return "possessive"
    if s[0] == " " and t[0].isupper() and t.replace("-","").replace(".","").isalpha():
        return "cap_word"
    if s[0] == " " and t[0].islower(): return "lower_word"
    if not t[0].isalnum(): return "punct"
    if t[0].isdigit(): return "digit"
    if t[0].islower(): return "glue_lower"
    if t[0].isupper(): return "glue_cap"
    return "other"

def entropy_bits(counter):
    n = sum(counter.values())
    if n == 0: return 0.0
    return -sum((c/n)*math.log2(c/n) for c in counter.values())

def gini(x):
    x = np.sort(np.asarray(x, dtype=np.float64))
    n = len(x)
    if n == 0 or x.sum() == 0: return 0.0
    return float((2*np.sum((np.arange(1,n+1))*x)/(n*x.sum())) - (n+1)/n)

def profile(lab, seqlen=1):
    """H1/H4 context profile for one label."""
    C = ctx_of(lab); P = pos_of(lab)
    n = len(P)
    if n == 0: return None
    prev1 = C[:, K-1]; prev2 = C[:, K-2]; nxt = C[:, K+seqlen]
    pc = Counter(prev_cat(p) for p in prev1)
    nc = Counter(next_cat(x) for x in nxt)
    prev_ctr = Counter(prev1.tolist()); next_ctr = Counter(nxt.tolist())
    bg_ctr = Counter(zip(prev2.tolist(), prev1.tolist()))
    # context markers in left window: girl/story cues
    marks = {" girl": 3828, " story": 3446, " Once": 9843}
    mark_ids = {}
    for w, i in marks.items():
        mark_ids[w] = i
    left = C[:, :K]
    girl_adj = np.zeros(n, bool)
    for w, i in mark_ids.items():
        girl_adj |= (left == i).any(axis=1)
    out = {
        "n": n,
        "prev_entropy_bits": round(entropy_bits(prev_ctr), 3),
        "prev_eff_vocab": round(2**entropy_bits(prev_ctr), 1),
        "prev_bigram_entropy_bits": round(entropy_bits(bg_ctr), 3),
        "next_entropy_bits": round(entropy_bits(next_ctr), 3),
        "next_eff_vocab": round(2**entropy_bits(next_ctr), 1),
        "prev_top1_share": round(prev_ctr.most_common(1)[0][1]/n, 4),
        "prev_top5_share": round(sum(c for _,c in prev_ctr.most_common(5))/n, 4),
        "next_top1_share": round(next_ctr.most_common(1)[0][1]/n, 4),
        "next_top5_share": round(sum(c for _,c in next_ctr.most_common(5))/n, 4),
        "prev_cats": {k: round(v/n, 4) for k, v in pc.most_common()},
        "next_cats": {k: round(v/n, 4) for k, v in nc.most_common()},
        "slot_cue_frac": round(sum(v for k, v in pc.items() if k in SLOT_CUES)/n, 4),
        "followed_by_cap_frac": round(nc.get("cap_word", 0)/n, 4),
        "standalone_frac": round(1 - (nc.get("cap_word",0)+nc.get("glue_cap",0)+nc.get("glue_lower",0)+nc.get("digit",0))/n, 4),
        "girlstory_marker_left16_frac": round(float(girl_adj.mean()), 4),
        "top20_prev": [[dec(i), c] for i, c in prev_ctr.most_common(20)],
        "top20_next": [[dec(i), c] for i, c in next_ctr.most_common(20)],
        "top15_prev_bigram": [[dec(a)+dec(b), c] for (a, b), c in bg_ctr.most_common(15)],
    }
    return out

def first_mention(pos_arr, union_pos=None):
    """H2: same-id (and optional any-variant) first-mention flags.
    Context for predicting token at g is [((g-1)//1024)*1024, g)."""
    p = np.sort(pos_arr)
    cstart = ((p - 1) // BLOCK) * BLOCK
    prev_same = np.concatenate(([np.int64(-1)], p[:-1]))
    fm_same = prev_same < cstart
    if union_pos is None:
        return p, fm_same, fm_same
    u = np.sort(union_pos)
    idx = np.searchsorted(u, p)          # u[idx-1] = nearest union occurrence strictly before p
    prev_any = np.where(idx > 0, u[np.maximum(idx - 1, 0)], -1)
    fm_any = prev_any < cstart
    return p, fm_same, fm_any

def dispersion(pos_arr):
    u = pos_arr // W - U0
    hist = np.bincount(u, minlength=U1-U0)
    doc_idx = np.searchsorted(eos_pos, pos_arr)
    dc = Counter(doc_idx.tolist())
    docs = np.array(sorted(dc.values()))[::-1]
    n = len(pos_arr)
    return {
        "n": int(n),
        "updates_with_ge1": int((hist > 0).sum()),
        "top_update_share": round(float(hist.max())/n, 4),
        "top10_updates_share": round(float(np.sort(hist)[-10:].sum())/n, 4),
        "gini_over_updates": round(gini(hist), 4),
        "n_docs": int(len(docs)),
        "sites_per_doc_mean": round(float(n/len(docs)), 2),
        "max_sites_one_doc": int(docs[0]),
        "top1_doc_share": round(float(docs[0])/n, 4),
        "top10_doc_share": round(float(docs[:10].sum())/n, 4),
        "docs_with_ge10_sites": int((docs >= 10).sum()),
        "frac_sites_in_docs_ge10": round(float(docs[docs >= 10].sum())/n, 4),
        "update_hist_argmax": int(np.argmax(hist) + U0),
    }, hist, dc

# ---------- run all ----------
results = {"profiles": {}, "first_mention": {}, "dispersion": {}}
hists = {}

# focus labels incl. controls
focus_labels = []
for nm in FOCUS:
    for form in ("sp", "bare"):
        lab = f"{form}_{nm}"
        if lab in labels:
            focus_labels.append(lab)

for lab in focus_labels:
    seqlen = len(h0["patterns"][lab]["ids"])
    prof = profile(lab, seqlen)
    if prof: results["profiles"][lab] = prof

# split controls: bare_Alice after-newline (was swapped in CF) vs not (survived, impotent)
cA = ctx_of("bare_Alice"); pA = pos_of("bare_Alice")
after_nl = cA[:, K-1] == NL
for sub, mask in (("bare_Alice_afterNL", after_nl), ("bare_Alice_other", ~after_nl)):
    if mask.sum() == 0: continue
    # build a temporary profile on the masked subset
    sub_ctx = cA[mask]
    prev_ctr = Counter(sub_ctx[:, K-1].tolist()); next_ctr = Counter(sub_ctx[:, K+1].tolist())
    pc = Counter(prev_cat(p) for p in sub_ctx[:, K-1])
    ncat = Counter(next_cat(x) for x in sub_ctx[:, K+1])
    nn = int(mask.sum())
    results["profiles"][sub] = {
        "n": nn,
        "prev_entropy_bits": round(entropy_bits(prev_ctr), 3),
        "next_entropy_bits": round(entropy_bits(next_ctr), 3),
        "prev_cats": {k: round(v/nn, 4) for k, v in pc.most_common()},
        "next_cats": {k: round(v/nn, 4) for k, v in ncat.most_common()},
        "slot_cue_frac": round(sum(v for k, v in pc.items() if k in SLOT_CUES)/nn, 4),
        "followed_by_cap_frac": round(ncat.get("cap_word", 0)/nn, 4),
        "top20_prev": [[dec(i), c] for i, c in prev_ctr.most_common(20)],
        "top20_next": [[dec(i), c] for i, c in next_ctr.most_common(20)],
    }

# megan_bare (line-initial "M"+"egan") — the H0 star: what follows/precedes
if "bare_Megan" in labels and len(pos_of("bare_Megan")):
    results["profiles"]["bare_Megan"] = profile("bare_Megan", seqlen=2)

# H2 + H3 for focus names (union variants for any-variant FM)
union_map = {}
for nm in FOCUS:
    parts = []
    for form in ("sp", "bare"):
        lab = f"{form}_{nm}"
        if lab in labels: parts.append(pos_of(lab))
    union_map[nm] = np.concatenate(parts) if parts else np.empty(0, np.int64)

for nm in FOCUS:
    lab = f"sp_{nm}"
    if lab not in labels or len(pos_of(lab)) == 0: continue
    p, fm_same, fm_any = first_mention(pos_of(lab), union_map[nm])
    # earlier-mentions-in-context histogram (same id)
    cstart = ((p - 1) // BLOCK) * BLOCK
    nprev = np.array([np.searchsorted(p, g) - np.searchsorted(p, cs)
                      for g, cs in zip(p, cstart)])
    results["first_mention"][nm] = {
        "n": int(len(p)),
        "first_mention_same_id": int(fm_same.sum()),
        "fm_same_frac": round(float(fm_same.mean()), 4),
        "first_mention_any_variant": int(fm_any.sum()),
        "fm_any_frac": round(float(fm_any.mean()), 4),
        "repeats": int((~fm_same).sum()),
        "earlier_mentions_hist": {str(k): int(v) for k, v in
                                  sorted(Counter(np.minimum(nprev, 5).tolist()).items())},
    }
    d, hist, dc = dispersion(pos_of(lab))
    results["dispersion"][nm] = d
    hists[nm] = hist

# bare_Megan dispersion (chat-log clustering test)
if "bare_Megan" in labels and len(pos_of("bare_Megan")):
    d, hist, dc = dispersion(pos_of("bare_Megan"))
    results["dispersion"]["Megan_bare_lineinitial"] = d
    hists["Megan_bare_lineinitial"] = hist

np.savez_compressed(os.path.join(STATS, "update_hists.npz"),
                    **{k: v for k, v in hists.items()})

# ---------- cross-name dose-response regression ----------
probe = {}
for r in map(json.loads, open(os.path.join(
        REPO, "out/olmo3_50b_ddp_resumecheck/name_slot_dist_v2.jsonl"))):
    if r["step"] in (50000, 55000) and r["prompt_label"] in ("girl_story", "girl_isnamed"):
        probe[(r["step"], r["prompt_label"])] = r["name_first_prob"]

reg_rows = []
for nm in GIRLS:
    lab = f"sp_{nm}"
    if lab not in labels: continue
    if len(h0["patterns"][lab]["ids"]) > 1:
        continue  # " Name" multi-token: probe P is a shared-first-token upper bound; excluded
    p = pos_of(lab)
    if len(p) == 0:
        raw = fm = slot = fm_slot = 0
        ndocs = 0; g = 0.0
    else:
        pp, fm_same, _ = first_mention(p)
        C = ctx_of(lab)
        cats = np.array([prev_cat(x) in SLOT_CUES for x in C[:, K-1]])
        # align cats (ctx rows are in ascending pos order == sorted order)
        raw = len(p); fm = int(fm_same.sum()); slot = int(cats.sum())
        order = np.argsort(p)
        fm_flags = fm_same  # first_mention() sorted internally; positions ascending already
        fm_slot = int((fm_flags & cats[order]).sum())
        d, _, _ = dispersion(p)
        ndocs = d["n_docs"]; g = d["gini_over_updates"]
    row = {"name": nm, "raw_sp": raw, "first_mention": fm, "slot_cue": slot,
           "fm_and_slot": fm_slot, "n_docs": ndocs, "gini": g}
    for (step, lblp), d_ in probe.items():
        v = d_.get(nm)
        row[f"P{step//1000}k_{'story' if lblp=='girl_story' else 'named'}"] = (
            v["p"] if isinstance(v, dict) else None)
    if row.get("P50k_story") is not None:
        row["dP_story"] = row["P55k_story"] - row["P50k_story"]
        row["dP_named"] = row["P55k_named"] - row["P50k_named"]
    reg_rows.append(row)

def spearman(x, y):
    try:
        from scipy.stats import spearmanr
        r, pv = spearmanr(x, y)
        return round(float(r), 3), float(pv)
    except ImportError:
        def rank(v):
            a = np.argsort(v); rk = np.empty(len(v)); rk[a] = np.arange(len(v))
            # average ties
            vv = np.asarray(v, float)
            for val in np.unique(vv):
                m = vv == val
                rk[m] = rk[m].mean()
            return rk
        rx, ry = rank(x), rank(y)
        r = float(np.corrcoef(rx, ry)[0, 1])
        return round(r, 3), float("nan")

reg = {"rows": sorted(reg_rows, key=lambda r: -(r.get("dP_story") or 0))}
have = [r for r in reg_rows if r.get("dP_story") is not None and r["raw_sp"] > 0]
try:
    xs = {k: [r[k] for r in have] for k in ("raw_sp","first_mention","slot_cue","fm_and_slot","n_docs")}
    y = [r["dP_story"] for r in have]
    y2 = [r["dP_named"] for r in have]
    reg["spearman_dP_story"] = {k: spearman(v, y) for k, v in xs.items()}
    reg["spearman_dP_named"] = {k: spearman(v, y2) for k, v in xs.items()}
    reg["n_names"] = len(have)
except Exception as e:
    reg["spearman_error"] = str(e)

with open(os.path.join(STATS, "analysis.json"), "w") as f:
    json.dump(results, f, indent=1)
with open(os.path.join(STATS, "regression.json"), "w") as f:
    json.dump(reg, f, indent=1)

# ---------- qualitative example dumps ----------
rng = random.Random(1234)
def dump_examples(lab, fname, n_rand=40):
    P = pos_of(lab); C = ctx_of(lab)
    if len(P) == 0: return
    lines = [f"# {lab}: {len(P)} sites; ctx +-{K} tokens; site token in [[...]]\n"]
    idxs = rng.sample(range(len(P)), min(n_rand, len(P)))
    seqlen = len(h0["patterns"][lab]["ids"])
    for i in sorted(idxs):
        row = C[i]
        left = tok.decode(row[:K].tolist()).replace("\n", "\\n")
        site = tok.decode(row[K:K+seqlen].tolist()).replace("\n", "\\n")
        right = tok.decode(row[K+seqlen:].tolist()).replace("\n", "\\n")
        u = int(P[i]) // W
        lines.append(f"upd {u} pos {int(P[i])}: {left}[[{site}]]{right}")
    # top-doc cluster examples
    doc_idx = np.searchsorted(eos_pos, P)
    top_docs = Counter(doc_idx.tolist()).most_common(3)
    for di, cnt in top_docs:
        lines.append(f"\n## top doc cluster: doc#{di} with {cnt} sites")
        sel = np.flatnonzero(doc_idx == di)[:5]
        for i in sel:
            row = C[i]
            left = tok.decode(row[:K].tolist()).replace("\n", "\\n")
            site = tok.decode(row[K:K+seqlen].tolist()).replace("\n", "\\n")
            right = tok.decode(row[K+seqlen:].tolist()).replace("\n", "\\n")
            lines.append(f"  {left}[[{site}]]{right}")
    with open(os.path.join(STATS, fname), "w") as f:
        f.write("\n".join(lines) + "\n")

for lab in focus_labels + (["bare_Megan"] if "bare_Megan" in labels else []):
    dump_examples(lab, f"examples_{lab}.txt")

print("analysis written to", STATS)
print(json.dumps({k: results["profiles"][k].get("slot_cue_frac") for k in results["profiles"]}, indent=1))