File size: 14,919 Bytes
327edac | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 | """Reconstruct per-agent, per-activation trajectories -- INCLUDING passive activations -- for
every run in divdata, and write `trajectories.parquet`.
Why this exists. `actions.parquet` only contains the two active actions the simulator logs
(create_post, create_comment). But every activation begins with a `refresh` row in `trace`
(the feed fetch) that carries a wall-clock timestamp whether or not the agent then acts, so
the *passive* side of the trajectory -- "agent was activated, looked at the feed, did
nothing" -- is fully recoverable from the database. Validated on the source runs: every
refresh window contains at most one action (1196/1200 windows have exactly one; the 4 with
zero are `status=exhausted` activations where all retries failed to parse).
Alignment to step. `step` is not in the database; it lives in `<run>_chats.jsonl`, one row
per activation. Step-0 rows have no refresh (the sign-up intro post is written directly), so
the k-th refresh of an agent is its k-th chats row with step > 0. Where that count matches,
step is EXACT; where an agent is off by one (seen in ~half the older sweeps) its steps are
NULL but its timestamps and actions are still emitted. Coverage is reported per run.
Reasoning. Some model families emit a free-text deliberation before the tool call; the
harness stored it in `<run>_reasoning.jsonl` keyed by (step, agent_id). It is joined on
where the step aligned. Absent for families that do not expose reasoning.
Decision record. Each activation's attempts in the chat sidecar carry the model's visible
output (`response_content` -- gemma and Magistral state their deliberation here ~95-99% of
the time, Qwen3 ~19%, gpt-oss/GLM-4 never), the tool call it chose (name + args, so the
target post is known even when the action later failed), and the parse error of each failed
attempt. All three are kept. The prompt is not: the feed it contained is `impressions`.
Row = one activation of one agent:
run_id, agent_id, model, activation_idx, step, refreshed_at, status, n_attempts,
action ('create_post' | 'create_comment' | 'none'), acted_at, latency_s,
post_id (own post for create_post; target post for create_comment), comment_id,
tool_name, tool_args (JSON), response_text, reasoning, finish_reason, errors,
extra_in_window (True for a 2nd+ action inside one refresh window; see below)
"""
import sqlite3, glob, os, json, argparse, collections, sys
import pandas as pd
ROOT = os.environ.get("OASIS_ROOT", ".") # directory holding runs/ (and attractor_rl/eval_runs*/)
sys.path.insert(0, os.path.dirname(__file__))
from build_divdata import agent_models
def decision(d):
"""The decision record of one activation, from its attempts list: the visible text the
model produced (its stated deliberation, where the family emits one), the tool call it
chose, and the parse errors of failed attempts. The prompt (user_message) is NOT kept --
the feed it contained is already in `impressions`, in slate order."""
atts = d.get("attempts") or []
last = next((a for a in reversed(atts) if "tool_calls" in a or "response_content" in a), None)
errs = [a.get("error") for a in atts if a.get("error")]
tc = (last or {}).get("tool_calls") or []
tc0 = tc[0] if tc else {}
return dict(n_attempts=len(atts),
response_text=((last or {}).get("response_content") or None),
finish_reason=((last or {}).get("finish_reason") or None),
tool_name=tc0.get("name"), tool_args=(json.dumps(tc0.get("args"), ensure_ascii=False) if tc0 else None),
errors=(" | ".join(str(e) for e in errs) or None))
def load_chats(path):
"""agent_id -> list of (step, status, decision) in file order, step>0 only; plus the
step-0 record per agent (the intro post, written without a refresh)."""
seq = collections.defaultdict(list); step0 = {}
for line in open(path):
try: d = json.loads(line)
except Exception: continue
try: st = int(d.get("step"))
except Exception: continue
ag = int(d.get("agent_id"))
if st == 0: step0[ag] = decision(d); continue
seq[ag].append((st, d.get("status"), decision(d)))
return seq, step0
def load_reasoning(path):
out = {}
if not os.path.exists(path): return out
for line in open(path):
try: d = json.loads(line)
except Exception: continue
try: out[(int(d["step"]), int(d["agent_id"]))] = d.get("reasoning")
except Exception: continue
return out
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--out", default="divdata")
ap.add_argument("--runs", default="divdata/runs.parquet")
ap.add_argument("--limit", type=int, default=0)
ap.add_argument("--workers", type=int, default=8)
ap.add_argument("--patch-steps", action="store_true",
help="fill NULL step in posts/comments/actions of --out from the trajectory alignment")
a = ap.parse_args()
runs = pd.read_parquet(a.runs, columns=["run_id", "source_path"])
if a.limit: runs = runs.head(a.limit)
T = []
log = dict(runs=0, no_chats=0, agents=0, agents_aligned=0, activations=0, passive=0,
windows_multi=0, agents_multi=0, with_reasoning=0, with_response_text=0, with_tool_call=0, status=collections.Counter())
from multiprocessing import Pool
with Pool(a.workers) as pool:
for i, (rows, plog) in enumerate(pool.imap_unordered(one_run, list(zip(runs.run_id, runs.source_path)), chunksize=4)):
T.extend(rows)
for k, v in plog.items():
if k == "status":
for kk, vv in v.items(): log["status"][kk] += vv
else: log[k] += v
if (i + 1) % 100 == 0:
print(f" ...{i+1}/{len(runs)} activations={len(T)}", flush=True)
finish(a, T, log)
def one_run(arg):
run_id, rel = arg
log = dict(runs=0, no_chats=0, agents=0, agents_aligned=0, activations=0, passive=0,
windows_multi=0, agents_multi=0, with_reasoning=0, with_response_text=0, with_tool_call=0, status=collections.Counter())
T = []
if True:
db = os.path.join(ROOT, rel)
meta_p = db.replace(".db", ".metadata.json")
_, am = agent_models(meta_p)
ch_p = db.replace(".db", "_chats.jsonl")
seq, step0 = load_chats(ch_p) if os.path.exists(ch_p) else (None, {})
if seq is None: log["no_chats"] += 1
reason = load_reasoning(db.replace(".db", "_reasoning.jsonl"))
try:
c = sqlite3.connect(db)
u2a = {u: g for u, g in c.execute("SELECT user_id, agent_id FROM user")}
# trace and post/comment timestamps differ by ~80 ms, so link by per-agent
# ordinal instead: the agent's k-th create_post in trace is its k-th post row.
posts_of = collections.defaultdict(list); coms_of = collections.defaultdict(list)
for pid, u in c.execute("SELECT post_id, user_id FROM post ORDER BY created_at, post_id"):
posts_of[u2a.get(u)].append(pid)
for cid, pid, u in c.execute("SELECT comment_id, post_id, user_id FROM comment ORDER BY created_at, comment_id"):
coms_of[u2a.get(u)].append((cid, pid))
tr = c.execute("SELECT user_id, created_at, action FROM trace WHERE action IN ('refresh','create_post','create_comment') "
"ORDER BY created_at, rowid").fetchall()
c.close()
except Exception:
return T, log
log["runs"] += 1
per = collections.defaultdict(list)
for u, ts, act in tr:
ag = u2a.get(u)
if ag is not None: per[ag].append((ts, act))
for ag, events in per.items():
model = am.get(ag, ("", None))[0] if am else None
s_ag = (seq or {}).get(ag, [])
n_ref = sum(1 for _, act in events if act == "refresh")
aligned = (seq is not None) and (len(s_ag) == n_ref)
log["agents"] += 1; log["agents_aligned"] += aligned
# walk windows
k = -1; cur = None; kp = kc = 0; multi = False
rows = []
for ts, act in events:
if act == "refresh":
if cur: rows.append(cur)
k += 1
st, status, dec = (s_ag[k] if aligned else (None, None, {}))
cur = dict(run_id=run_id, agent_id=ag, model=model, activation_idx=k, step=st,
refreshed_at=ts, status=status, action="none",
acted_at=None, post_id=None, comment_id=None, **dec,
reasoning=(reason.get((st, ag)) if st is not None else None))
else:
if cur is None:
# step-0 intro post: no refresh precedes it
rows.append(dict(run_id=run_id, agent_id=ag, model=model, activation_idx=-1, step=0,
refreshed_at=None, status="ok", action=act, **step0.get(ag, {}),
acted_at=ts, post_id=(posts_of[ag][kp] if kp < len(posts_of[ag]) else None),
comment_id=None, reasoning=reason.get((0, ag))))
kp += 1
continue
if cur["action"] != "none":
# A second action inside one refresh window. 16 runs (one old v9
# hetero_late archive run accounts for 2,468 of 3,247 such actions)
# violate the one-action-per-activation assumption, so the refresh
# alignment is not trustworthy for this agent: the extra action gets
# its own row (same activation_idx, extra_in_window=True) and the
# whole agent's step/decision fields are nulled below.
log["windows_multi"] += 1; multi = True
rows.append(cur)
cur = dict(cur, action="none", acted_at=None, post_id=None, comment_id=None, extra_in_window=True)
cur["action"] = act; cur["acted_at"] = ts
if act == "create_post":
cur["post_id"] = posts_of[ag][kp] if kp < len(posts_of[ag]) else None; kp += 1
else:
cid, pid = coms_of[ag][kc] if kc < len(coms_of[ag]) else (None, None); kc += 1
cur["comment_id"] = cid; cur["post_id"] = pid
if cur: rows.append(cur)
if multi:
log["agents_multi"] += 1
for r in rows:
if r["activation_idx"] >= 0:
for f in ("step", "status", "n_attempts", "response_text", "finish_reason",
"tool_name", "tool_args", "errors", "reasoning"):
r[f] = None
for r in rows:
r.setdefault("extra_in_window", False)
log["activations"] += 1
log["passive"] += (r["action"] == "none")
log["with_reasoning"] += (r["reasoning"] is not None)
log["with_response_text"] += bool(r.get("response_text"))
log["with_tool_call"] += bool(r.get("tool_name"))
log["status"][r["status"]] += 1
T.extend(rows)
return T, log
def finish(a, T, log):
df = pd.DataFrame(T)
# older runs wrote second-precision timestamps, newer ones microseconds
df["refreshed_at"] = pd.to_datetime(df.refreshed_at, format="mixed"); df["acted_at"] = pd.to_datetime(df.acted_at, format="mixed")
df["latency_s"] = (df.acted_at - df.refreshed_at).dt.total_seconds().round(3)
df["step"] = df.step.astype("Int64"); df["n_attempts"] = df.n_attempts.astype("Int64")
for col in ("response_text", "finish_reason", "tool_name", "tool_args", "errors", "reasoning"):
if col not in df: df[col] = None
df["post_id"] = df.post_id.astype("Int64"); df["comment_id"] = df.comment_id.astype("Int64")
df = df[["run_id", "agent_id", "model", "activation_idx", "step", "refreshed_at", "status", "n_attempts",
"action", "acted_at", "latency_s", "post_id", "comment_id",
"tool_name", "tool_args", "response_text", "reasoning", "finish_reason", "errors", "extra_in_window"]]
p = os.path.join(a.out, "trajectories.parquet")
df.to_parquet(p, index=False, compression="zstd")
log["status"] = dict(log["status"])
log["rows"] = len(df); log["size_mb"] = round(os.path.getsize(p) / 1e6, 1)
log["step_coverage"] = round(float(df.step.notna().mean()), 4)
log["action_counts"] = df.action.value_counts().to_dict()
print(json.dumps(log, indent=1, default=str))
if a.patch_steps:
patch_steps(a.out, df, log)
json.dump(log, open(os.path.join(a.out, "_trajectories_build_log.json"), "w"), indent=1, default=str)
def patch_steps(out, t, log):
"""The refresh-window alignment holds for every agent that has a chat sidecar, whereas
build_divdata's action-count alignment fails for an agent whose ok-activation and
action counts differ. Where the two agree they agree exactly (verified 100% on every
comparable row), so use the window alignment to fill step where build_divdata left NULL.
Existing non-null values are never overwritten."""
key = t[t.action != "none"].dropna(subset=["step"])
for name, on, sub in (("posts", ["run_id", "post_id"], key[key.action == "create_post"]),
("comments", ["run_id", "comment_id"], key[key.action == "create_comment"]),
("actions", ["run_id", "agent_id", "created_at"], key.assign(created_at=key.acted_at))):
p = os.path.join(out, f"{name}.parquet")
if not os.path.exists(p): continue
df = pd.read_parquet(p)
if name == "actions": df["created_at"] = pd.to_datetime(df.created_at, format="mixed")
m = df.merge(sub[on + ["step"]].rename(columns={"step": "_s"}), on=on, how="left")
before = df.step.isna().sum()
both = m.dropna(subset=["step", "_s"])
agree = float((both.step.astype(int) == both._s.astype(int)).mean()) if len(both) else 1.0
m["step"] = m.step.where(m.step.notna(), m._s).astype("Int64")
m = m.drop(columns="_s")
if name == "actions": m["created_at"] = m.created_at.astype(str)
m.to_parquet(p, index=False, compression="zstd")
log[f"patch_{name}"] = dict(null_before=int(before), null_after=int(m.step.isna().sum()),
rows=len(m), agreement_on_overlap=round(agree, 5))
print(f" patched {name}: step NULL {before} -> {int(m.step.isna().sum())} of {len(m)}; overlap agreement {agree:.5f}")
if __name__ == "__main__":
main()
|