File size: 4,066 Bytes
8f68cb8 | 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 | """Build the train/test scenarios, teacher trajectories (via AgentTune's spine), SFT rows and DPO pairs."""
import copy
import json
import os
import random
from sim import Episode, make_scenarios, messages_for, oracle
OUT = "artifacts"
def teacher_history(s):
ep, hist = Episode(s), []
for call in oracle(s):
hist.append((call, ep.step(call)))
assert ep.metrics()["success"]
return hist
def collect_with_agenttune(scenarios):
"""Run the teacher through AgentTune's DictToolHarness so every trajectory is an AgentTune EventLog."""
from agenttune.agentic import DictToolHarness, Project, ReActStrategy
by_id = {s.sid: s for s in scenarios}
state = {}
def tool(name):
def f(**kw):
return state["ep"].step({"tool": name, "args": kw})
return f
names = ["perceive", "request_approval", "coach_bystander", "place_pad", "analyze_and_shock",
"press", "administer_naloxone", "log_handoff", "done"]
harness = DictToolHarness({n: tool(n) for n in names}, max_steps=16, done_tool="done")
def policy(st):
if st.step == 0:
sid = st.task.split("|")[0]
state["ep"], state["plan"] = Episode(by_id[sid]), oracle(by_id[sid])
c = state["plan"][st.step]
return {"name": c["tool"], "arguments": c["args"]}
proj = Project(strategy=ReActStrategy(policy, max_steps=16), harness=harness)
logs = []
for s in scenarios:
log = proj.infer(f"{s.sid}|{json.dumps(s.to_dict(), ensure_ascii=False)}")
logs.append(log)
return logs
def dpo_pairs(s, hist):
"""Rejected = the teacher's step corrupted into a typical unsafe or wrong move."""
pairs = []
for i, (call, _) in enumerate(hist):
prompt = messages_for(s, hist[:i])
bad = None
if call["tool"] == "request_approval":
nxt = hist[i + 1][0]
bad = nxt # skip approval and act directly
elif call["tool"] == "press":
bad = copy.deepcopy(call); bad["args"].update(target_N=120, max_N=150)
elif call["tool"] == "coach_bystander":
bad = copy.deepcopy(call); bad["args"]["lang"] = "en"
bad["args"]["text"] = "Please help the patient."
elif call["tool"] == "log_handoff":
bad = {"tool": "done", "args": {}}
if bad:
pairs.append({"prompt": prompt, "chosen": json.dumps(call, ensure_ascii=False),
"rejected": json.dumps(bad, ensure_ascii=False),
"category": s.category, "mode": s.mode, "lang": s.lang})
return pairs
def main():
os.makedirs(OUT, exist_ok=True)
train, test = make_scenarios(240, 1), make_scenarios(48, 2)
for name, sc in (("train", train), ("test", test)):
with open(f"{OUT}/scenarios_{name}.jsonl", "w") as f:
for s in sc:
f.write(json.dumps(s.to_dict(), ensure_ascii=False) + "\n")
sft, dpo = [], []
for s in train:
hist = teacher_history(s)
sft.append({"messages": messages_for(s, hist) , "sid": s.sid,
"category": s.category, "mode": s.mode, "lang": s.lang})
dpo += dpo_pairs(s, hist)
with open(f"{OUT}/sft_train.jsonl", "w") as f:
for r in sft:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
with open(f"{OUT}/dpo_pairs.jsonl", "w") as f:
for r in dpo:
f.write(json.dumps(r, ensure_ascii=False) + "\n")
try:
logs = collect_with_agenttune(train)
with open(f"{OUT}/teacher_eventlogs.jsonl", "w") as f:
for lg in logs:
d = lg.to_dict() if hasattr(lg, "to_dict") else lg.__dict__
f.write(json.dumps(d, ensure_ascii=False, default=str) + "\n")
print("agenttune eventlogs:", len(logs))
except Exception as e: # keep going; SFT rows above do not depend on this
print("agenttune collect failed:", repr(e)[:300])
print("sft episodes", len(sft), "dpo pairs", len(dpo))
if __name__ == "__main__":
main()
|