File size: 10,192 Bytes
0584798 | 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 | from __future__ import annotations
import argparse
import json
import sys
from pathlib import Path
BASE_DIR = Path(__file__).resolve().parent.parent
if str(BASE_DIR) not in sys.path:
sys.path.insert(0, str(BASE_DIR))
from combined_inference import classify_query
from config import BASE_DIR, INTENT_HEAD_CONFIG, ensure_artifact_dirs
from model_runtime import get_head
from schemas import validate_classify_response
DEFAULT_THRESHOLDS = [0.0, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45]
SAFE_INTENT_TYPES = {"ambiguous", "personal_reflection", "support"}
OBVIOUS_INTENT_TYPES = {"informational", "commercial", "transactional"}
SWEEP_SUITE_PATH = BASE_DIR / "examples" / "intent_threshold_sweep_suite.json"
OUTPUT_PATH = BASE_DIR / "artifacts" / "evaluation" / "intent_threshold_sweep.json"
def load_jsonl(path: Path) -> list[dict]:
with path.open("r", encoding="utf-8") as handle:
return [json.loads(line) for line in handle]
def load_json(path: Path) -> list[dict]:
return json.loads(path.read_text(encoding="utf-8"))
def round_score(value: float) -> float:
return round(float(value), 4)
def write_json(path: Path, payload: dict) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n", encoding="utf-8")
def evaluate_intent_head_threshold(threshold: float) -> dict:
head = get_head("intent_type")
dataset_specs = [
("val", INTENT_HEAD_CONFIG.split_paths["val"]),
("test", INTENT_HEAD_CONFIG.split_paths["test"]),
("hard_cases", BASE_DIR / "data" / "hard_cases.jsonl"),
("third_wave_cases", BASE_DIR / "data" / "third_wave_cases.jsonl"),
]
rows = []
for suite_name, path in dataset_specs:
for item in load_jsonl(path):
rows.append({"suite": suite_name, **item})
predictions = head.predict_batch([row["text"] for row in rows], confidence_threshold=threshold)
obvious_total = 0
obvious_fallback = 0
ambiguous_total = 0
ambiguous_bad_allow = 0
intent_only_safe_pred = 0
for row, prediction in zip(rows, predictions):
predicted_label = prediction["label"]
head_would_fallback = (not prediction["meets_confidence_threshold"]) or (predicted_label in SAFE_INTENT_TYPES)
if row[INTENT_HEAD_CONFIG.label_field] in OBVIOUS_INTENT_TYPES:
obvious_total += 1
if head_would_fallback:
obvious_fallback += 1
if row[INTENT_HEAD_CONFIG.label_field] == "ambiguous":
ambiguous_total += 1
if not head_would_fallback:
ambiguous_bad_allow += 1
if head_would_fallback and predicted_label in SAFE_INTENT_TYPES:
intent_only_safe_pred += 1
return {
"obvious_prompt_count": obvious_total,
"obvious_false_fallback_rate": round_score(obvious_fallback / obvious_total) if obvious_total else 0.0,
"ambiguous_prompt_count": ambiguous_total,
"ambiguous_bad_allow_rate": round_score(ambiguous_bad_allow / ambiguous_total) if ambiguous_total else 0.0,
"safe_predicate_rate": round_score(intent_only_safe_pred / len(rows)) if rows else 0.0,
}
def evaluate_combined_threshold(threshold: float) -> dict:
suite = load_json(SWEEP_SUITE_PATH)
benchmark = load_json(BASE_DIR / "examples" / "demo_prompt_suite.json")
obvious_total = 0
obvious_false_fallback = 0
safe_total = 0
safe_bad_allow = 0
intent_only = 0
phase_only = 0
both = 0
policy_safe = 0
suite_outputs = []
for item in suite:
payload = validate_classify_response(classify_query(item["input"], threshold_overrides={"intent_type": threshold}))
fallback = payload["model_output"].get("fallback")
fallback_applied = fallback is not None
failed_components = set((fallback or {}).get("failed_components", []))
if item["expected_outcome"] == "pass":
obvious_total += 1
if fallback_applied:
obvious_false_fallback += 1
else:
safe_total += 1
if not fallback_applied:
safe_bad_allow += 1
if fallback_applied:
if failed_components == {"intent_type"}:
intent_only += 1
elif failed_components == {"decision_phase"}:
phase_only += 1
elif failed_components == {"intent_type", "decision_phase"}:
both += 1
else:
policy_safe += 1
suite_outputs.append(
{
"input": item["input"],
"expected_outcome": item["expected_outcome"],
"fallback_applied": fallback_applied,
"failed_components": sorted(failed_components),
"intent_type": payload["model_output"]["classification"]["intent"]["type"],
"decision_phase": payload["model_output"]["classification"]["intent"]["decision_phase"],
"intent_confidence": payload["model_output"]["classification"]["intent"]["component_confidence"]["intent_type"]["confidence"],
"phase_confidence": payload["model_output"]["classification"]["intent"]["component_confidence"]["decision_phase"]["confidence"],
}
)
benchmark_fallbacks = []
for item in benchmark:
payload = validate_classify_response(classify_query(item["input"], threshold_overrides={"intent_type": threshold}))
fallback = payload["model_output"].get("fallback")
fallback_applied = fallback is not None
failed_components = set((fallback or {}).get("failed_components", []))
benchmark_fallbacks.append(
{
"fallback_applied": fallback_applied,
"intent_only": failed_components == {"intent_type"},
"phase_only": failed_components == {"decision_phase"},
"both": failed_components == {"intent_type", "decision_phase"},
}
)
total_suite_fallbacks = intent_only + phase_only + both + policy_safe
benchmark_total_fallbacks = sum(1 for item in benchmark_fallbacks if item["fallback_applied"])
return {
"suite_path": str(SWEEP_SUITE_PATH),
"obvious_prompt_count": obvious_total,
"false_fallback_rate_on_obvious_prompts": round_score(obvious_false_fallback / obvious_total) if obvious_total else 0.0,
"safe_prompt_count": safe_total,
"bad_allow_rate_on_safe_prompts": round_score(safe_bad_allow / safe_total) if safe_total else 0.0,
"fallback_responsibility": {
"intent_only": intent_only,
"phase_only": phase_only,
"both": both,
"policy_safe": policy_safe,
"intent_share_of_threshold_fallbacks": round_score(
(intent_only + both) / (intent_only + phase_only + both)
)
if (intent_only + phase_only + both)
else 0.0,
"phase_share_of_threshold_fallbacks": round_score(
(phase_only + both) / (intent_only + phase_only + both)
)
if (intent_only + phase_only + both)
else 0.0,
"fallback_rate": round_score(total_suite_fallbacks / len(suite)) if suite else 0.0,
},
"benchmark_fallback_rate": round_score(benchmark_total_fallbacks / len(benchmark)) if benchmark else 0.0,
"benchmark_intent_only_fallback_rate": round_score(
sum(1 for item in benchmark_fallbacks if item["intent_only"]) / len(benchmark)
)
if benchmark
else 0.0,
"benchmark_phase_only_fallback_rate": round_score(
sum(1 for item in benchmark_fallbacks if item["phase_only"]) / len(benchmark)
)
if benchmark
else 0.0,
"suite_outputs": suite_outputs,
}
def pick_recommended_threshold(results: list[dict]) -> dict:
return min(
results,
key=lambda item: (
item["combined"]["bad_allow_rate_on_safe_prompts"],
item["head"]["ambiguous_bad_allow_rate"],
item["combined"]["false_fallback_rate_on_obvious_prompts"],
abs(item["combined"]["fallback_responsibility"]["intent_share_of_threshold_fallbacks"] - 0.5),
item["combined"]["benchmark_fallback_rate"],
item["threshold"],
),
)
def apply_threshold(threshold: float) -> None:
calibration_path = INTENT_HEAD_CONFIG.calibration_path
payload = json.loads(calibration_path.read_text(encoding="utf-8"))
payload["confidence_threshold"] = round_score(threshold)
payload["threshold_selection_mode"] = "manual_sweep"
write_json(calibration_path, payload)
def main() -> None:
parser = argparse.ArgumentParser(description="Sweep candidate intent_type thresholds and compare end-to-end behavior.")
parser.add_argument(
"--thresholds",
nargs="*",
type=float,
default=DEFAULT_THRESHOLDS,
help="Candidate thresholds to evaluate.",
)
parser.add_argument(
"--apply-threshold",
type=float,
default=None,
help="Optional threshold to write into the intent_type calibration artifact after evaluation.",
)
args = parser.parse_args()
ensure_artifact_dirs()
thresholds = [round_score(value) for value in args.thresholds]
results = []
for threshold in thresholds:
results.append(
{
"threshold": threshold,
"head": evaluate_intent_head_threshold(threshold),
"combined": evaluate_combined_threshold(threshold),
}
)
recommended = pick_recommended_threshold(results)
output = {
"thresholds": thresholds,
"results": results,
"recommended_threshold": recommended["threshold"],
}
write_json(OUTPUT_PATH, output)
if args.apply_threshold is not None:
apply_threshold(args.apply_threshold)
output["applied_threshold"] = round_score(args.apply_threshold)
write_json(OUTPUT_PATH, output)
print(json.dumps(output, indent=2))
if __name__ == "__main__":
main()
|