Spaces:
Sleeping
Sleeping
File size: 8,059 Bytes
d02bacd | 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 | """
Build a small SFT dataset for the Chief of Staff from oracle rollouts.
Output:
- ``training/sft_data/cos_sft.jsonl`` for chat/message based trainers.
- ``training/sft_data/cos_sft_autotrain.jsonl`` for AutoTrain SFT, with a
single ``text`` column.
Format (Hugging Face chat-template friendly):
{"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "<observation snapshot>"},
{"role": "assistant", "content": "<JSON action>"}
]}
No GPU and no API key required. Run before SFT:
python3 training/build_sft_dataset.py
"""
from __future__ import annotations
import json
import sys
from pathlib import Path
from typing import Any
REPO = Path(__file__).resolve().parents[2]
if str(REPO) not in sys.path:
sys.path.insert(0, str(REPO))
from ceo_brief_env.environment import ( # noqa: E402
CEOBriefEnvironment,
oracle_action_for_observation,
required_experts_for_task,
)
from ceo_brief_env.models import CoSAction, CoSObservation # noqa: E402
OUT_DIR = REPO / "training" / "sft_data"
OUT_PATH = OUT_DIR / "cos_sft.jsonl"
AUTOTRAIN_OUT_PATH = OUT_DIR / "cos_sft_autotrain.jsonl"
TASKS = ["easy_brief", "medium_brief", "hard_brief", "expert_brief"]
RAG_MODES = [False, True]
SYSTEM_PROMPT = (
"You are the Chief of Staff (CoS) in AutoDataLab++. You orchestrate four "
"specialists: analyst, finance, strategy, hr. At each step, decide ONE "
"action and reply with STRICT JSON only.\n\n"
"Schema: {\"action_type\": one of [\"consult\", \"ask\", \"summarize\", \"submit\", \"noop\"], "
"\"expert_id\": one of [\"analyst\", \"finance\", \"hr\", \"strategy\"] or null, "
"\"sub_question_id\": string or null, \"notes\": string or null}.\n"
"Rules: consult each required expert at most once before summarizing; "
"summarize before submitting; submit only when the brief is composed."
)
SYSTEM_PROMPTS = [
SYSTEM_PROMPT,
(
"Act as AutoDataLab++ Chief of Staff. Choose the next orchestration step. "
"Return strict JSON only with keys action_type, expert_id, sub_question_id, notes. "
"Valid actions: consult, ask, summarize, submit, noop. Valid experts: analyst, finance, hr, strategy."
),
(
"You are a routing policy for a CEO brief environment. Consult missing required experts first, "
"then summarize, then submit. Output one JSON object only; no markdown and no explanation."
),
]
OBSERVATION_VARIANTS = range(6)
def render_observation(obs: CoSObservation, variant: int = 0) -> str:
required = required_experts_for_task(obs.task_name)
missing = [e for e in required if e not in obs.consulted_experts]
parts = [
f"task_name: {obs.task_name}",
f"task_difficulty: {obs.task_difficulty}",
f"step_count: {obs.step_count}",
f"max_steps: {obs.max_steps}",
f"rag_enabled: {obs.rag_enabled}",
f"available_experts: {obs.available_experts}",
f"required_experts: {required}",
f"consulted_experts: {obs.consulted_experts}",
f"missing_required_experts: {missing}",
f"current_brief_present: {obs.current_brief is not None}",
f"data_quality_score: {round(float(obs.data_quality_score or 0.0), 4)}",
f"recent_issues: {obs.issues[-3:]}",
f"instruction: {obs.instruction}",
]
if variant == 0:
return "\n".join(parts)
if variant == 1:
return (
f"Task={obs.task_name}; difficulty={obs.task_difficulty}; "
f"step={obs.step_count}/{obs.max_steps}; rag={obs.rag_enabled}; "
f"required={required}; consulted={obs.consulted_experts}; missing={missing}; "
f"brief_ready={obs.current_brief is not None}; dq={round(float(obs.data_quality_score or 0.0), 4)}"
)
if variant == 2:
return json.dumps(
{
"task": obs.task_name,
"difficulty": obs.task_difficulty,
"step": obs.step_count,
"max_steps": obs.max_steps,
"rag_enabled": obs.rag_enabled,
"required_experts": required,
"consulted_experts": obs.consulted_experts,
"missing_required_experts": missing,
"brief_ready": obs.current_brief is not None,
"data_quality_score": round(float(obs.data_quality_score or 0.0), 4),
},
separators=(",", ":"),
)
if variant == 3:
return (
"Decision checklist:\n"
f"- task: {obs.task_name}\n"
f"- required experts: {', '.join(required)}\n"
f"- already consulted: {', '.join(obs.consulted_experts) or 'none'}\n"
f"- still missing: {', '.join(missing) or 'none'}\n"
f"- brief composed: {obs.current_brief is not None}\n"
f"- choose the single next action"
)
if variant == 4:
if missing:
status = f"The next priority is one of the missing experts: {missing}."
elif obs.current_brief is None:
status = "All required experts are consulted, but the brief is not summarized yet."
else:
status = "The brief is summarized and ready for final submission."
return (
f"You are at step {obs.step_count} for {obs.task_name}. "
f"Consulted experts are {obs.consulted_experts}. {status} "
f"RAG mode is {obs.rag_enabled}."
)
return (
f"state: task={obs.task_name} | required={required} | consulted={obs.consulted_experts} | "
f"missing={missing} | has_brief={obs.current_brief is not None} | action?"
)
def action_to_json(action: CoSAction) -> str:
payload: dict[str, Any] = action.model_dump(exclude_none=True)
return json.dumps(payload, separators=(",", ":"), sort_keys=True)
def collect_records() -> list[dict[str, Any]]:
records: list[dict[str, Any]] = []
for task in TASKS:
for use_rag in RAG_MODES:
env = CEOBriefEnvironment()
obs = env.reset(task=task, use_rag=use_rag)
while not obs.done and obs.step_count < obs.max_steps:
action = oracle_action_for_observation(obs)
assistant_msg = action_to_json(action)
for system_prompt in SYSTEM_PROMPTS:
for variant in OBSERVATION_VARIANTS:
user_msg = render_observation(obs, variant=variant)
records.append(
{
"messages": [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_msg},
{"role": "assistant", "content": assistant_msg},
]
}
)
obs = env.step(action)
return records
def render_autotrain_text(record: dict[str, Any]) -> str:
messages = record["messages"]
system = messages[0]["content"]
user = messages[1]["content"]
assistant = messages[2]["content"]
return (
"<|im_start|>system\n"
f"{system}<|im_end|>\n"
"<|im_start|>user\n"
f"{user}<|im_end|>\n"
"<|im_start|>assistant\n"
f"{assistant}<|im_end|>"
)
def main() -> int:
OUT_DIR.mkdir(parents=True, exist_ok=True)
records = collect_records()
with OUT_PATH.open("w", encoding="utf-8") as f:
for rec in records:
f.write(json.dumps(rec, ensure_ascii=False) + "\n")
with AUTOTRAIN_OUT_PATH.open("w", encoding="utf-8") as f:
for rec in records:
f.write(json.dumps({"text": render_autotrain_text(rec)}, ensure_ascii=False) + "\n")
print(f"[sft-data] wrote {len(records)} examples to {OUT_PATH}")
print(f"[sft-data] wrote AutoTrain text dataset to {AUTOTRAIN_OUT_PATH}")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|