Spaces:
Running
Running
File size: 11,222 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 | """Headline benchmark with the eval-mode safety net OFF.
Compares four routing policies on the three hard tasks, with five seeds and
both RAG settings. The safety net is OFF (``auto_fill_required=False``) so an
incomplete brief gets penalised by the grader.
Policies
--------
- ``base_naive``: untrained baseline, consults the analyst then submits.
- ``base_roundrobin``: untrained second baseline, walks experts in fixed order.
- ``trained_mlp``: actually trained CoS policy. A 2-layer MLP routing policy
(REINFORCE, 600 episodes, lr 0.003) loaded from
``training/checkpoints/cos_final.pt``. This is the headline trained-model
number. If torch / the checkpoint isn't available the row is skipped and the
plot annotates that.
- ``oracle_router``: deterministic upper bound. The handcoded routing policy
that always consults the required experts in the canonical order. We label
this an *upper bound*, not a trained model -- our trained policies (the MLP
above and the SFT/GRPO LLMs) are trained to imitate this behaviour, and the
number tells you how high a perfect routing policy can score on the
current grader / RAG settings.
Outputs
-------
- ``training/evidence/headline_benchmark.json`` -- raw cells + per-seed runs.
- ``training/evidence/plots/headline_terminal_reward.{png,svg}`` -- the chart.
"""
from __future__ import annotations
import json
import random
import statistics
import sys
from pathlib import Path
from typing import Any, Callable
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,
)
from ceo_brief_env.models import CoSAction, CoSObservation # noqa: E402
HARD_TASKS = ["hard_brief", "expert_brief", "crisis_brief"]
SEEDS = [11, 23, 47, 91, 137]
RAG_SETTINGS = [False, True]
def naive_picker(obs: CoSObservation) -> CoSAction:
"""Untrained base policy: consult analyst, then submit. Will be incomplete."""
if "analyst" not in obs.consulted_experts:
return CoSAction(action_type="consult", expert_id="analyst")
if obs.current_brief is None:
return CoSAction(action_type="summarize")
return CoSAction(action_type="submit")
def roundrobin_picker(obs: CoSObservation) -> CoSAction:
"""Second base policy: walks experts in fixed order then submits."""
for expert in ["finance", "analyst", "hr", "strategy"]:
if expert not in obs.consulted_experts:
return CoSAction(action_type="consult", expert_id=expert)
if obs.current_brief is None:
return CoSAction(action_type="summarize")
return CoSAction(action_type="submit")
def _load_trained_mlp_picker() -> tuple[Callable[[CoSObservation], CoSAction], str] | None:
"""Load the actually-trained MLP CoS routing policy.
Returns ``(picker_fn, info)`` or ``None`` if torch / the checkpoint isn't
available. The MLP was trained with REINFORCE for 600 episodes (see
``training/scripts/train_cos_local.py``); ``cos_final.pt`` is the resulting
state dict.
"""
try:
import numpy as np # noqa: F401
import torch
sys.path.insert(0, str(REPO / "training" / "scripts"))
from train_cos_local import ( # type: ignore
ACTIONS,
PolicyNet,
featurize,
load_policy_state_dict_from_file,
)
except Exception as exc: # pragma: no cover -- env without torch
return None
ckpt_paths = [
REPO / "training" / "checkpoints" / "cos_final.pt",
REPO / "training" / "checkpoints" / "cos_ckpt0.pt",
]
ckpt = next((p for p in ckpt_paths if p.exists()), None)
if ckpt is None:
return None
model = PolicyNet()
info = load_policy_state_dict_from_file(model, ckpt)
model.eval()
def picker(obs: CoSObservation) -> CoSAction:
feats = torch.from_numpy(featurize(obs)).unsqueeze(0)
with torch.no_grad():
logits = model(feats)
idx = int(torch.argmax(logits, dim=-1).item())
return ACTIONS[idx]
return picker, f"{ckpt.name} ({info})"
def build_policies() -> tuple[dict[str, Callable[[CoSObservation], CoSAction]], dict[str, str]]:
"""Return the policy table plus a metadata dict for the plot legend."""
policies: dict[str, Callable[[CoSObservation], CoSAction]] = {
"base_naive": naive_picker,
"base_roundrobin": roundrobin_picker,
}
info: dict[str, str] = {
"base_naive": "untrained baseline (analyst-only)",
"base_roundrobin": "untrained baseline (fixed order)",
}
trained = _load_trained_mlp_picker()
if trained is not None:
policies["trained_mlp"] = trained[0]
info["trained_mlp"] = f"trained MLP CoS (REINFORCE, 600 ep) - {trained[1]}"
policies["oracle_router"] = oracle_action_for_observation
info["oracle_router"] = "oracle router (upper bound, handcoded canonical sequence)"
return policies, info
def run_episode(picker: Callable[[CoSObservation], CoSAction], task: str, use_rag: bool, seed: int) -> dict[str, Any]:
random.seed(seed)
env = CEOBriefEnvironment(shaping="strict", auto_fill_required=False)
obs = env.reset(task=task, use_rag=use_rag)
max_steps = obs.max_steps
steps = 0
consulted_path: list[str] = []
while not obs.done and steps < max_steps:
steps += 1
action = picker(obs)
if action.action_type == "consult" and action.expert_id:
consulted_path.append(action.expert_id)
obs = env.step(action)
terminal = float(obs.terminal_grader_score or 0.0)
cumulative = float(obs.reward_breakdown.cumulative if obs.reward_breakdown else 0.0)
return {
"terminal": terminal,
"cumulative": cumulative,
"steps": steps,
"consulted": list(obs.consulted_experts),
"path": consulted_path,
"submitted": bool(obs.done),
}
def aggregate(samples: list[float]) -> dict[str, float]:
if not samples:
return {"mean": 0.0, "std": 0.0, "n": 0}
if len(samples) == 1:
return {"mean": samples[0], "std": 0.0, "n": 1}
return {
"mean": statistics.fmean(samples),
"std": statistics.pstdev(samples),
"n": len(samples),
}
def main() -> None:
out_dir = REPO / "training" / "evidence"
out_dir.mkdir(parents=True, exist_ok=True)
plots_dir = out_dir / "plots"
plots_dir.mkdir(parents=True, exist_ok=True)
policies, policy_info = build_policies()
results: dict[str, Any] = {
"schema": "autodatalab-plus.headline_benchmark.v2",
"config": {
"tasks": HARD_TASKS,
"policies": list(policies.keys()),
"policy_info": policy_info,
"seeds": SEEDS,
"rag_settings": RAG_SETTINGS,
"auto_fill_required": False,
"shaping": "strict",
},
"cells": [],
}
for task in HARD_TASKS:
for use_rag in RAG_SETTINGS:
for policy_name, policy_fn in policies.items():
terminals: list[float] = []
cumulatives: list[float] = []
runs: list[dict[str, Any]] = []
for seed in SEEDS:
rollout = run_episode(policy_fn, task, use_rag, seed)
terminals.append(rollout["terminal"])
cumulatives.append(rollout["cumulative"])
runs.append({"seed": seed, **rollout})
cell = {
"task": task,
"policy": policy_name,
"use_rag": use_rag,
"terminal": aggregate(terminals),
"cumulative": aggregate(cumulatives),
"runs": runs,
}
results["cells"].append(cell)
print(
f"task={task:13s} rag={str(use_rag):5s} policy={policy_name:18s} "
f"terminal_mean={cell['terminal']['mean']:.3f} "
f"std={cell['terminal']['std']:.3f}"
)
json_path = out_dir / "headline_benchmark.json"
json_path.write_text(json.dumps(results, indent=2))
print(f"\nwrote {json_path.relative_to(REPO)}")
try:
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
except ImportError:
print("matplotlib not available; skipping money plot")
return
by_cell: dict[tuple[str, str], list[float]] = {}
for cell in results["cells"]:
key = (cell["task"], cell["policy"])
by_cell.setdefault(key, []).extend([r["terminal"] for r in cell["runs"]])
plot_order = [p for p in ("base_naive", "base_roundrobin", "trained_mlp", "oracle_router") if p in policies]
pretty_name = {
"base_naive": "base (naive)",
"base_roundrobin": "base (round-robin)",
"trained_mlp": "trained MLP CoS\n(REINFORCE, 600 ep)",
"oracle_router": "oracle router\n(upper bound)",
}
color = {
"base_naive": "#9aa0a6",
"base_roundrobin": "#bdc1c6",
"trained_mlp": "#34a853",
"oracle_router": "#1a73e8",
}
hatch = {
"oracle_router": "//",
}
fig, ax = plt.subplots(figsize=(10.4, 5.0), dpi=140)
n_groups = len(HARD_TASKS)
width = 0.78 / max(1, len(plot_order))
xs = list(range(n_groups))
for i, policy in enumerate(plot_order):
means = []
stds = []
for task in HARD_TASKS:
samples = by_cell.get((task, policy), [])
agg = aggregate(samples)
means.append(agg["mean"])
stds.append(agg["std"])
offsets = [x + (i - (len(plot_order) - 1) / 2.0) * width for x in xs]
bars = ax.bar(
offsets,
means,
width=width,
yerr=stds,
capsize=4,
color=color[policy],
edgecolor="black",
linewidth=0.6,
hatch=hatch.get(policy, ""),
label=pretty_name[policy],
)
for bar, mean in zip(bars, means):
ax.text(
bar.get_x() + bar.get_width() / 2,
bar.get_height() + 0.015,
f"{mean:.2f}",
ha="center",
va="bottom",
fontsize=8,
)
ax.set_xticks(xs)
ax.set_xticklabels([t.replace("_brief", "") for t in HARD_TASKS])
ax.set_ylim(0.0, 1.0)
ax.set_ylabel("Terminal grader score (0..1)")
ax.set_title(
"Terminal reward, fallback disabled\n"
"untrained baselines vs trained MLP CoS vs oracle router (upper bound)\n"
"3 hard tasks, 5 seeds (RAG on/off averaged)"
)
ax.grid(axis="y", linestyle="--", alpha=0.4)
ax.legend(loc="upper left", framealpha=0.9, fontsize=9)
fig.tight_layout()
png_path = plots_dir / "headline_terminal_reward.png"
svg_path = plots_dir / "headline_terminal_reward.svg"
fig.savefig(png_path)
fig.savefig(svg_path)
print(f"wrote {png_path.relative_to(REPO)}")
print(f"wrote {svg_path.relative_to(REPO)}")
if __name__ == "__main__":
main()
|