File size: 26,365 Bytes
c3b572e | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 | #!/usr/bin/env python3
"""CE v19 RustPPM Hermes instruction-agent evaluation lane.
This ports the v18 bounded bridge action-selection gate onto a pure RustPPM token
stream so v19 reports tool accuracy separately from phoneme PPL. It includes a
small symbolic intent feature layer so the PPM can generalize across Hermes-style
paraphrases instead of only memorizing exact probe strings.
"""
from __future__ import annotations
import json
import math
import os
import re
import sys
import time
from pathlib import Path
from typing import Sequence
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
import ce_ppm
# obs, goal, expected action, result text
V19_TOOL_PROBES = [
("question asks arithmetic twelve times thirteen", "answer accurately", "CALCULATE", "one hundred fifty six"),
("question asks current public fact", "answer accurately", "SEARCH", "search before answering"),
("question asks prior conversation fact", "answer accurately", "MEMORY_LOOKUP", "retrieve memory before answering"),
("question lacks required user preference", "answer accurately", "ASK_USER", "ask a concise clarification"),
("training log shows ppl increasing and validation degrading", "steer training", "STEER_TRAINING", "stop bad trajectory and launch corrected run"),
("tool output has an error traceback", "debug systematically", "DEBUG", "inspect root cause patch test relaunch"),
("code change needs verification", "verify before claiming done", "VERIFY", "verified against live output"),
("file needs inspection before answering", "read first then answer", "READ_FILE", "file contents read"),
("file needs creation or update", "write then verify", "WRITE_FILE", "file written and verified"),
("tests need to be run", "run and report real output", "RUN_TESTS", "tests executed"),
]
# Extra training paraphrases. They intentionally vary surface wording while
# sharing latent Hermes action semantics via `_hermes_instruction_features`.
V19_HERMES_TRAIN_PROBES = V19_TOOL_PROBES + [
("calculate 12 multiplied by 13", "answer with the number", "CALCULATE", "156"),
("what is the latest release news online", "use current sources", "SEARCH", "web search first"),
("what did we decide in the previous session", "use stored context", "MEMORY_LOOKUP", "memory recalled"),
("I have not provided the target environment", "avoid guessing", "ASK_USER", "ask for target"),
("validation loss is climbing and the run is going bad", "fix training trajectory", "STEER_TRAINING", "cancel or patch run"),
("pytest returned a traceback in the terminal", "find root cause", "DEBUG", "inspect failure"),
("before saying done confirm the patch works", "verify real output", "VERIFY", "verified"),
("open the config and inspect contents", "read first", "READ_FILE", "file read"),
("create the missing script file", "write artifact", "WRITE_FILE", "file written"),
("run the regression suite", "execute tests", "RUN_TESTS", "tests run"),
("multiply twenty one and six", "compute exactly", "CALCULATE", "126"),
("look up the current API docs", "get current facts", "SEARCH", "docs searched"),
("recall our saved preference for this project", "retrieve memory", "MEMORY_LOOKUP", "preference recalled"),
("ambiguous request missing repo name", "clarify missing information", "ASK_USER", "clarifying question"),
("HF job PPL got worse after step 500", "intervene in training", "STEER_TRAINING", "training steered"),
("command failed with ModuleNotFoundError", "debug systematically", "DEBUG", "module root cause found"),
("show evidence before claiming landed", "verify before response", "VERIFY", "evidence checked"),
("read LOGS.md before answering", "inspect file", "READ_FILE", "logs read"),
("update LOGS.md with the metric", "write update", "WRITE_FILE", "log updated"),
("execute pytest for the touched tests", "run tests", "RUN_TESTS", "pytest executed"),
]
V19_HELDOUT_HERMES_PROBES = [
("what's 37 plus 58", "compute exactly", "CALCULATE", "95"),
("check the current HuggingFace job status online", "use live source", "SEARCH", "status searched"),
("use memory to remember what checkpoint we chose", "retrieve prior fact", "MEMORY_LOOKUP", "checkpoint recalled"),
("not enough detail to choose staging or prod", "ask instead of guessing", "ASK_USER", "asked target"),
("the training curve is diverging after launch", "actively repair the run", "STEER_TRAINING", "bad run corrected"),
("the stack trace points at a failing import", "root cause debug", "DEBUG", "debugged import"),
("prove the code works before reporting success", "verify with command output", "VERIFY", "proof collected"),
("inspect the JSON results file", "read the artifact", "READ_FILE", "json read"),
("save this new evaluator script", "write file", "WRITE_FILE", "script saved"),
("run the unit tests now", "execute validation", "RUN_TESTS", "unit tests run"),
]
V19_LATEST_TURN_PROBES = [
("history: what is 12 times 13. latest user: read LOGS.md before replying", "inspect latest request", "READ_FILE", "logs read"),
("history: read file. latest user: run pytest now", "validate latest request", "RUN_TESTS", "tests run"),
("history: run tests. latest user: what is 2 plus 2", "answer latest request", "CALCULATE", "4"),
("history: current docs needed. latest user: missing repo name", "handle latest ambiguity", "ASK_USER", "asked repo"),
]
V19_POLICY_ORDER_PROBES = [
("Please update the file after reading it first", "perform prerequisite order", "READ_FILE", "file read before patch"),
("Open /tmp/a.py and write /tmp/b.py based on it", "read before write", "READ_FILE", "source file read"),
("Search the web for current docs, then edit config.yaml", "perform all needed steps", "SEARCH", "docs searched before edit"),
("Use memory if needed, but current public fact asks latest release", "answer accurately", "SEARCH", "current source searched"),
("Tests failed after my patch; inspect the traceback before editing", "debug before patch", "DEBUG", "root cause found"),
]
V19_SWEBENCH_PROBES = [
("SWE-bench issue: reproduce failing pytest, inspect src/package/core.py, patch it, then run pytest tests/test_core.py", "solve repository bug with evidence", "RUN_TESTS", "repro test run first"),
("Failure log points to src/lib/parser.py; read the file before changing anything", "inspect source before patch", "READ_FILE", "source inspected"),
("Traceback from pytest shows AttributeError in package/module.py", "debug root cause before editing", "DEBUG", "root cause isolated"),
("Patch src/lib/parser.py to fix the regression", "modify code artifact", "WRITE_FILE", "patch applied"),
("After patch, run pytest tests/test_parser.py -q and report the real output", "validate patch", "RUN_TESTS", "target tests run"),
("Before saying resolved, verify the failing SWE-bench test and no regression", "prove fix", "VERIFY", "verification complete"),
]
_INTENT_RULES = [
("INTENT_CALCULATE", r"\b(calculate|multiply|multiplied|plus|minus|divide|compute|arithmetic|\d+\s*(\+|\*|x|/|-))\b"),
("INTENT_SEARCH", r"\b(current|latest|online|web|search|look up|docs|status online|live source|public fact)\b"),
("INTENT_MEMORY", r"\b(memory|remember|recall|previous|prior|stored|saved preference|checkpoint we chose)\b"),
("INTENT_ASK_USER", r"\b(ambiguous|missing|not enough|clarify|lacks|required|provided|target environment|staging or prod)\b"),
("INTENT_STEER_TRAINING", r"\b(training|ppl|validation|loss|curve|diverging|degrading|HF job|run is going bad|launch)\b"),
("INTENT_DEBUG", r"\b(traceback|error|failed|failing|debug|root cause|ModuleNotFoundError|stack trace)\b"),
("INTENT_VERIFY", r"\b(verify|prove|evidence|confirm|before claiming|before reporting|works|landed)\b"),
("INTENT_READ_FILE", r"\b(read|inspect|open|contents|LOGS\.md|JSON results|artifact)\b"),
("INTENT_WRITE_FILE", r"\b(write|create|update|save|missing script|file written|LOGS\.md)\b"),
("INTENT_RUN_TESTS", r"\b(test|tests|pytest|unit tests|regression suite|validation suite)\b"),
]
def _latest_instruction_text(text: str) -> str:
"""Return the latest user/request segment from a chat transcript-like string."""
markers = ["latest user:", "[new message]", "new message:", "user:"]
lowered = text.lower()
cut = -1
marker_len = 0
for marker in markers:
idx = lowered.rfind(marker)
if idx > cut:
cut = idx
marker_len = len(marker)
return text[cut + marker_len :].strip() if cut >= 0 else text
def _hermes_instruction_features(obs: str, goal: str) -> str:
latest_obs = _latest_instruction_text(obs)
# Intent features are based on the latest user/request text, not generic
# controller goals like "answer latest request" that can contain misleading
# words (e.g. "latest" should not imply SEARCH).
text = latest_obs.lower()
feats = ["HERMES_INSTRUCTION"]
for name, pattern in _INTENT_RULES:
if re.search(pattern, text):
feats.append(name)
# Preserve one-shot induction identity in the same symbolic feature channel.
# Without this, all "copy demonstrated mapping" probes share the same suffix
# before ACTION and a high-order PPM correctly ties the candidates.
sym = re.search(r"\bsymbol\s+(\d+)\b", text)
if sym:
feats.append(f"SYMBOL_{sym.group(1)}")
return " ".join(feats)
def _encode_obs(obs: str, goal: str) -> str:
# Put stable intent atoms first so high-order suffixes include semantics even
# when the natural-language paraphrase differs.
return f"{_hermes_instruction_features(obs, goal)} USER_TEXT {obs}"
def _encode_goal(obs: str, goal: str) -> str:
# Repeat stable intent atoms at the end of GOAL so the fixed "choose action"
# thought suffix still leaves semantic intent inside the PPM context window.
return f"GOAL_TEXT {goal} {_hermes_instruction_features(obs, goal)}"
def _unique_action_labels(probes: Sequence[tuple[str, str, str, str]]) -> list[str]:
labels = []
for _obs, _goal, action, _result in probes:
label = f"ACT_{action}"
if label not in labels:
labels.append(label)
return labels
def _candidate_labels() -> list[str]:
return _unique_action_labels(
V19_HERMES_TRAIN_PROBES
+ V19_HELDOUT_HERMES_PROBES
+ V19_LATEST_TURN_PROBES
+ V19_POLICY_ORDER_PROBES
+ V19_SWEBENCH_PROBES
)
def _action_id(tok, action_label: str) -> int:
return int(tok.action_token(action_label.replace("ACT_", "")))
_ACTION_DECISION_SPECS = {
"ACT_RESPOND": {"tool": "none", "needs_tool": False, "needs_user": False, "policy": "answer_in_english"},
"ACT_CALCULATE": {"tool": "calculator", "needs_tool": True, "needs_user": False, "policy": "compute_exact"},
"ACT_SEARCH": {"tool": "web_search", "needs_tool": True, "needs_user": False, "policy": "fetch_current_source"},
"ACT_MEMORY_LOOKUP": {"tool": "memory_search", "needs_tool": True, "needs_user": False, "policy": "retrieve_persistent_context"},
"ACT_ASK_USER": {"tool": "clarify", "needs_tool": False, "needs_user": True, "policy": "ask_missing_required_context"},
"ACT_STEER_TRAINING": {"tool": "training_control", "needs_tool": True, "needs_user": False, "policy": "intervene_on_bad_trajectory"},
"ACT_DEBUG": {"tool": "debugger", "needs_tool": True, "needs_user": False, "policy": "root_cause_then_patch"},
"ACT_VERIFY": {"tool": "verification_command", "needs_tool": True, "needs_user": False, "policy": "prove_before_claim"},
"ACT_READ_FILE": {"tool": "read_file", "needs_tool": True, "needs_user": False, "policy": "inspect_source_artifact"},
"ACT_WRITE_FILE": {"tool": "write_or_patch_file", "needs_tool": True, "needs_user": False, "policy": "modify_artifact_then_verify"},
"ACT_RUN_TESTS": {"tool": "terminal_tests", "needs_tool": True, "needs_user": False, "policy": "run_relevant_tests"},
}
_LEAK_SENTINELS = ("[fabric]", "[sessions]", "[qdrant]", "[facts]", "system prompt", "developer message", "MEMORY.md", "USER.md")
_ALLOWED_DECISION_KEYS = {"action", "tool", "needs_tool", "needs_user", "policy", "confidence", "args"}
def _contains_leak(value) -> bool:
text = value if isinstance(value, str) else json.dumps(value, sort_keys=True, default=str)
low = text.lower()
return any(s.lower() in low for s in _LEAK_SENTINELS)
def _redact_leaks(text: str) -> str:
out = str(text)
for sentinel in _LEAK_SENTINELS:
out = re.sub(re.escape(sentinel), "[REDACTED]", out, flags=re.IGNORECASE)
return out
def _first_path(text: str) -> str | None:
m = re.search(r"(/[^\s,:;]+|[A-Za-z0-9_./-]+\.(?:py|json|md|txt|toml|ya?ml))", text)
return m.group(1).rstrip(".,;:") if m else None
def _extract_command(text: str) -> str:
m = re.search(r"(pytest(?:\s+[-\w./=]+)*)", text)
return m.group(1).strip() if m else "pytest -q"
def _safe_args(args: dict) -> bool:
if not isinstance(args, dict):
return False
for key, value in args.items():
if not isinstance(key, str):
return False
if not isinstance(value, (str, int, float, bool, type(None))):
return False
if _contains_leak(value):
return False
return True
def _decision_args(action_label: str, obs: str, goal: str) -> dict:
latest = _redact_leaks(_latest_instruction_text(obs))
if _contains_leak(latest):
latest = "[REDACTED]"
path = _first_path(latest) or ""
if action_label == "ACT_CALCULATE":
return {"expression": latest[:120]}
if action_label == "ACT_SEARCH":
return {"query": latest[:160], "freshness_required": True}
if action_label == "ACT_MEMORY_LOOKUP":
return {"query": latest[:160]}
if action_label == "ACT_ASK_USER":
return {"question": "What missing target or constraint should I use?"}
if action_label == "ACT_STEER_TRAINING":
return {"signal": latest[:160], "action": "inspect_or_intervene"}
if action_label == "ACT_DEBUG":
return {"error": latest[:160], "method": "root_cause_first"}
if action_label == "ACT_VERIFY":
return {"command": _extract_command(latest), "evidence_target": "real command output"}
if action_label == "ACT_READ_FILE":
return {"path": path or "<path-from-latest-request>"}
if action_label == "ACT_WRITE_FILE":
return {"path": path or "<path-from-latest-request>", "mode": "patch"}
if action_label == "ACT_RUN_TESTS":
return {"command": _extract_command(latest)}
return {}
def render_decision_json(action_label: str, obs: str, goal: str) -> str:
"""Render the chosen action as a strict Hermes instruction-agent decision."""
if action_label not in _ACTION_DECISION_SPECS:
raise ValueError(f"unknown action label: {action_label}")
spec = _ACTION_DECISION_SPECS[action_label]
decision = {
"action": action_label,
"tool": spec["tool"],
"needs_tool": bool(spec["needs_tool"]),
"needs_user": bool(spec["needs_user"]),
"policy": spec["policy"],
"confidence": 1.0,
"args": _decision_args(action_label, obs, goal),
}
return json.dumps(decision, sort_keys=True, separators=(",", ":"))
def _args_valid_for_action(action: str, args: dict) -> bool:
if not _safe_args(args):
return False
required = {
"ACT_CALCULATE": ("expression",),
"ACT_SEARCH": ("query", "freshness_required"),
"ACT_MEMORY_LOOKUP": ("query",),
"ACT_ASK_USER": ("question",),
"ACT_STEER_TRAINING": ("signal", "action"),
"ACT_DEBUG": ("error", "method"),
"ACT_VERIFY": ("command", "evidence_target"),
"ACT_READ_FILE": ("path",),
"ACT_WRITE_FILE": ("path", "mode"),
"ACT_RUN_TESTS": ("command",),
}.get(action, ())
return all(k in args and args[k] not in ("", None) for k in required)
def validate_decision_json(text: str, expected_action: str) -> dict:
try:
payload = json.loads(text)
except Exception as exc:
return {"valid_json": False, "correct_action": False, "schema_valid": False, "known_action": False, "spec_valid": False, "confidence_valid": False, "args_valid": False, "leak_free": False, "tool_call_valid": False, "error": type(exc).__name__}
if not isinstance(payload, dict):
return {"valid_json": True, "correct_action": False, "schema_valid": False, "known_action": False, "spec_valid": False, "confidence_valid": False, "args_valid": False, "leak_free": False, "tool_call_valid": False, "payload": payload}
action = payload.get("action")
spec = _ACTION_DECISION_SPECS.get(action)
schema_valid = set(payload.keys()) == _ALLOWED_DECISION_KEYS
known_action = spec is not None
correct_action = action == expected_action
confidence = payload.get("confidence")
confidence_valid = isinstance(confidence, (int, float)) and not isinstance(confidence, bool) and 0.0 <= float(confidence) <= 1.0
spec_valid = bool(
known_action
and payload.get("tool") == spec["tool"]
and payload.get("needs_tool") == bool(spec["needs_tool"])
and payload.get("needs_user") == bool(spec["needs_user"])
and payload.get("policy") == spec["policy"]
)
args_valid = _args_valid_for_action(str(action), payload.get("args"))
leak_free = not _contains_leak(payload)
tool_call_valid = bool(schema_valid and known_action and correct_action and spec_valid and confidence_valid and args_valid and leak_free)
return {
"valid_json": True,
"correct_action": bool(correct_action),
"schema_valid": bool(schema_valid),
"known_action": bool(known_action),
"spec_valid": bool(spec_valid),
"confidence_valid": bool(confidence_valid),
"args_valid": bool(args_valid),
"leak_free": bool(leak_free),
"tool_call_valid": bool(tool_call_valid),
"payload": payload,
}
def train_tool_policy(
*,
repeats: int = 60,
max_order: int = 12,
probes=V19_HERMES_TRAIN_PROBES + V19_POLICY_ORDER_PROBES + V19_SWEBENCH_PROBES,
):
"""Train RustPPM on curated v19 Hermes action episodes."""
tok = ce_ppm.RustTokenizer()
ppm = ce_ppm.RustPPM(max_order, 1e-4, 0.25, 0.0)
for label in _candidate_labels():
_action_id(tok, label)
for _ in range(int(repeats)):
for obs, goal, action, result in probes:
seq = tok.typed_episode(_encode_obs(obs, goal), _encode_goal(obs, goal), action, result, 1.0)
ppm.update_sequence(seq)
return ppm, tok
def choose_action(ppm, tok, obs: str, goal: str, candidates: Sequence[str]) -> dict:
"""Choose the candidate action with highest PPM probability after policy prefix."""
for c in candidates:
_action_id(tok, c)
prefix = tok.policy_prefix(_encode_obs(obs, goal), _encode_goal(obs, goal))
vocab_size = int(len(tok))
scored = []
for cand in candidates:
aid = _action_id(tok, cand)
p, order, mass = ppm.prob_next(prefix, aid, vocab_size)
scored.append({"action": cand, "p": float(p), "logp": math.log(max(float(p), 1e-12)), "order": int(order), "mass": float(mass)})
scored.sort(key=lambda x: x["logp"], reverse=True)
return {"action": scored[0]["action"], "scores": scored, "valid": scored[0]["action"] in set(candidates)}
def _agentic_ppl(ppm, tok, probes: Sequence[tuple[str, str, str, str]]) -> float:
nlls = []
for obs, goal, action, result in probes:
seq = tok.typed_episode(_encode_obs(obs, goal), _encode_goal(obs, goal), action, result, 1.0)
nlls.append(float(ppm.sequence_nll(seq, len(tok))))
return math.exp(sum(nlls) / max(len(nlls), 1))
def _induction_eval(*, max_order: int = 12) -> dict:
"""One-shot action mapping induction on fresh symbols."""
tok = ce_ppm.RustTokenizer()
ppm = ce_ppm.RustPPM(max_order, 1e-4, 0.25, 0.0)
labels = ["ACT_ALPHA", "ACT_BETA", "ACT_GAMMA", "ACT_DELTA"]
for label in labels:
_action_id(tok, label)
for i, label in enumerate(labels):
obs = f"symbol {i}"
goal = "copy demonstrated mapping"
seq = tok.typed_episode(_encode_obs(obs, goal), _encode_goal(obs, goal), label.replace("ACT_", ""), "correct", 1.0)
ppm.update_sequence(seq)
correct = 0
details = []
for i, expected in enumerate(labels):
out = choose_action(ppm, tok, f"symbol {i}", "copy demonstrated mapping", labels)
correct += int(out["action"] == expected)
details.append({"symbol": i, "expected": expected, "predicted": out["action"], "top_scores": out["scores"][:3]})
return {"induction_accuracy": correct / len(labels), "induction_details": details}
def evaluate_tool_policy(ppm, tok, probes=V19_TOOL_PROBES, *, run_induction: bool = True) -> dict:
candidates = _candidate_labels()
correct = 0
valid = 0
details = []
t0 = time.perf_counter()
latencies = []
for obs, goal, expected_action, _result in probes:
start = time.perf_counter()
out = choose_action(ppm, tok, obs, goal, candidates)
latencies.append((time.perf_counter() - start) * 1000.0)
expected = f"ACT_{expected_action}"
decision_json = render_decision_json(out["action"], obs, goal)
decision_check = validate_decision_json(decision_json, expected)
correct += int(out["action"] == expected)
valid += int(bool(out["valid"]))
details.append({
"obs": obs,
"features": _hermes_instruction_features(obs, goal),
"expected": expected,
"predicted": out["action"],
"valid": out["valid"],
"decision_json": decision_json,
"decision_check": decision_check,
"top_scores": out["scores"][:5],
})
decision_checks = [d["decision_check"] for d in details]
metrics = {
"tool_selection_accuracy": correct / max(len(probes), 1),
"valid_action_accuracy": valid / max(len(probes), 1),
"decision_json_validity": sum(int(c["valid_json"] and c["schema_valid"]) for c in decision_checks) / max(len(decision_checks), 1),
"decision_schema_accuracy": sum(int(c["schema_valid"]) for c in decision_checks) / max(len(decision_checks), 1),
"decision_action_accuracy": sum(int(c["correct_action"]) for c in decision_checks) / max(len(decision_checks), 1),
"decision_leak_free_rate": sum(int(c["leak_free"]) for c in decision_checks) / max(len(decision_checks), 1),
"strict_tool_call_accuracy": sum(int(c["tool_call_valid"]) for c in decision_checks) / max(len(decision_checks), 1),
"agentic_ppl": _agentic_ppl(ppm, tok, probes),
"latency_ms_mean": sum(latencies) / max(len(latencies), 1),
"latency_ms_max": max(latencies) if latencies else 0.0,
"vocab_size": int(len(tok)),
"ppm_tables": int(ppm.table_count()),
"eval_wall_s": time.perf_counter() - t0,
"details": details,
}
if run_induction:
metrics.update(_induction_eval())
return metrics
def run() -> dict:
repeats = int(os.environ.get("CE_V19_TOOL_REPEATS", "60"))
max_order = int(os.environ.get("CE_V19_TOOL_MAX_ORDER", "12"))
ppm, tok = train_tool_policy(repeats=repeats, max_order=max_order)
train_metrics = evaluate_tool_policy(ppm, tok, V19_TOOL_PROBES, run_induction=True)
heldout_metrics = evaluate_tool_policy(ppm, tok, V19_HELDOUT_HERMES_PROBES, run_induction=False)
latest_turn_metrics = evaluate_tool_policy(ppm, tok, V19_LATEST_TURN_PROBES, run_induction=False)
policy_order_metrics = evaluate_tool_policy(ppm, tok, V19_POLICY_ORDER_PROBES, run_induction=False)
swebench_metrics = evaluate_tool_policy(ppm, tok, V19_SWEBENCH_PROBES, run_induction=False)
metrics = dict(train_metrics)
metrics["heldout_hermes"] = heldout_metrics
metrics["latest_turn"] = latest_turn_metrics
metrics["policy_order"] = policy_order_metrics
metrics["swebench_instruction"] = swebench_metrics
gate_sets = [train_metrics, heldout_metrics, latest_turn_metrics, policy_order_metrics, swebench_metrics]
metrics.update({
"status": "PASS" if all(m["tool_selection_accuracy"] >= 0.95 and m["valid_action_accuracy"] == 1.0 and m["decision_json_validity"] == 1.0 and m["decision_leak_free_rate"] == 1.0 and m["strict_tool_call_accuracy"] == 1.0 for m in gate_sets) and train_metrics["induction_accuracy"] >= 0.95 else "FAIL",
"repeats": repeats,
"max_order": max_order,
"cpu_only": True,
})
out = ROOT / "language_pipeline" / "training_results_v19_tool_eval.json"
out.write_text(json.dumps(metrics, indent=2), encoding="utf-8")
print(
f"V19_TOOL_EVAL status={metrics['status']} tool_acc={train_metrics['tool_selection_accuracy']:.3f} "
f"heldout_tool_acc={heldout_metrics['tool_selection_accuracy']:.3f} "
f"latest_turn_acc={latest_turn_metrics['tool_selection_accuracy']:.3f} "
f"policy_order_acc={policy_order_metrics['tool_selection_accuracy']:.3f} "
f"swebench_instr_acc={swebench_metrics['tool_selection_accuracy']:.3f} "
f"strict_tool_call={train_metrics['strict_tool_call_accuracy']:.3f}/{heldout_metrics['strict_tool_call_accuracy']:.3f}/{swebench_metrics['strict_tool_call_accuracy']:.3f} "
f"valid_action_acc={train_metrics['valid_action_accuracy']:.3f}/{heldout_metrics['valid_action_accuracy']:.3f} "
f"json_valid={train_metrics['decision_json_validity']:.3f}/{heldout_metrics['decision_json_validity']:.3f} "
f"leak_free={train_metrics['decision_leak_free_rate']:.3f}/{heldout_metrics['decision_leak_free_rate']:.3f} "
f"ind_acc={train_metrics['induction_accuracy']:.3f} agentic_ppl={train_metrics['agentic_ppl']:.3f} "
f"heldout_agentic_ppl={heldout_metrics['agentic_ppl']:.3f} lat_max={max(train_metrics['latency_ms_max'], heldout_metrics['latency_ms_max']):.3f}ms "
f"vocab={metrics['vocab_size']} tables={metrics['ppm_tables']} results={out}",
flush=True,
)
return metrics
if __name__ == "__main__":
run()
|