verify_star.py

import json, collections, random
from dataset import expand_data
from stokenizer import STokenizer
from graph_metrics import _distances, _classify

tok = STokenizer()
d = json.load(open("data/star_2arm_L6_valid_coconut.json"))
L = 6

def reach(edges, src):
    adj = collections.defaultdict(list)
    for a, b in edges: adj[a].append(b)
    seen, q = {src}, [src]
    while q:
        u = q.pop()
        for v in adj[u]:
            if v not in seen: seen.add(v); q.append(v)
    return seen

bad = 0
for s in d:
    r = reach(s["edges"], s["root"])
    # target reachable, neg_target NOT reachable, unique path depth L
    if s["target"] not in r: bad += 1; continue
    if s["neg_target"] in r: bad += 1; continue
    fdist, bdist, Ld = _distances(s["edges"], s["root"], s["target"])
    if Ld != L: bad += 1; continue
    # neighbor_k path must be a real chain root->target on the target arm
    if any(s["neighbor_k"][str(k)][0] not in fdist or fdist[s["neighbor_k"][str(k)][0]] != k for k in range(1, L+1)):
        bad += 1
print(f"structural check: {len(d)-bad}/{len(d)} valid (target reachable, neg unreachable, depth=={L}, unique path)")

# tokenization check on one sample, all hops + final answer
s = d[0]
print("\nsample root/target/neg:", s["root"], s["target"], s["neg_target"])
for k in range(1, L + 2):  # 1..L intermediate, L+1 = [A] answer
    q, cont = expand_data(s, k, len(s["steps"]))
    ids_q = tok.encode(q, add_special_tokens=False)
    ids_c = tok.encode(cont, add_special_tokens=False)
    n_lat = q.count("<|latent|>")
    tag = "[A]answer" if k == L + 1 else f"hop{k}"
    print(f"  {tag:9s}: {n_lat} latents -> target '{cont}'  (q_tokens={len(ids_q)}, ok)")

# category-metric sanity: decoy arm inflates frontier vs optimal
fdist, bdist, Ld = _distances(s["edges"], s["root"], s["target"])
frontier_nodes = [g for g in fdist if fdist[g] == 1]
print("\nhop-1 frontier (both arms' first nodes):", frontier_nodes,
      "| optimal (target arm only):", [g for g in frontier_nodes if _classify(g,1,fdist,bdist,Ld)[2]])