87 lines
4.1 KiB
Python
Executable File
87 lines
4.1 KiB
Python
Executable File
#!/usr/bin/env python3
|
|
"""Run only the bundled fictional cases against the private batch endpoint."""
|
|
|
|
import argparse
|
|
import hashlib
|
|
import json
|
|
from pathlib import Path
|
|
import sys
|
|
|
|
sys.path.insert(0, str(Path(__file__).resolve().parents[1]))
|
|
from hermes_batch_client import BatchClient, MODELS, canonical
|
|
from contracts import FAMILY, PROFILE, validate_families, validate_profile
|
|
|
|
|
|
ROOT = Path(__file__).resolve().parent
|
|
|
|
|
|
def source_cases():
|
|
"""Load built-in synthetic records, never a workbook or a user-selected file."""
|
|
fixture = json.loads((ROOT / "synthetic_cases.json").read_text())
|
|
assert fixture["synthetic"] is True and fixture["campaign_id"] == "SYNTHETIC"
|
|
return [{"campaign_id": fixture["campaign_id"], "suite_id": fixture["suite_id"], **case}
|
|
for case in fixture["cases"]]
|
|
|
|
|
|
def envelope(model, mode, sources, thinking):
|
|
"""Build versioned synthetic input with explicit sampling and context budgets."""
|
|
if mode == "transport":
|
|
prompt = 'Return only JSON with the field "status" equal to "LOCAL_OK".'
|
|
schema = {"type": "object", "properties": {"status": {"type": "string", "enum": ["LOCAL_OK"]}},
|
|
"required": ["status"], "additionalProperties": False}
|
|
count, context = 128, 16384
|
|
else:
|
|
name = "profile" if mode == "profile" else "family"
|
|
prompt = (ROOT / (name + "_prompt.txt")).read_text()
|
|
prompt += "\nOutput schema:\n" + json.dumps(PROFILE if mode == "profile" else FAMILY)
|
|
prompt += "\nSynthetic source records:\n" + json.dumps(sources, ensure_ascii=False)
|
|
schema = PROFILE if mode == "profile" else FAMILY
|
|
count, context = (2048, 16384) if mode == "profile" else (8192, 32768)
|
|
return {"campaign_id": "SYNTHETIC", "suite_id": "SYN-SUITE-01",
|
|
"source_sha256": hashlib.sha256(canonical(sources)).hexdigest(),
|
|
"prompt_version": "synthetic-" + mode + "-v1",
|
|
"request": {"model": model, "prompt": prompt, "stream": False, "think": thinking,
|
|
"format": schema, "options": {
|
|
"num_ctx": context, "num_predict": count,
|
|
"temperature": 0.6 if thinking else 0, "top_p": 0.95 if thinking else 1,
|
|
"top_k": 20, "seed": 42}}}
|
|
|
|
|
|
def main():
|
|
"""Emit synthetic timing and contract results without displaying source text."""
|
|
parser = argparse.ArgumentParser(description=__doc__)
|
|
parser.add_argument("--token-file", required=True)
|
|
parser.add_argument("--cache-dir", required=True)
|
|
parser.add_argument("--mode", choices=("transport", "profile", "group"), default="transport")
|
|
parser.add_argument("--model", choices=list(MODELS), required=True)
|
|
parser.add_argument("--case-id", default="SYN-010")
|
|
parser.add_argument("--think", action="store_true")
|
|
args = parser.parse_args()
|
|
sources = source_cases()
|
|
if args.mode == "profile":
|
|
sources = [source for source in sources if source["case_id"] == args.case_id]
|
|
if len(sources) != 1:
|
|
parser.error("case-id must select one bundled synthetic case")
|
|
request = envelope(args.model, args.mode, sources, args.think)
|
|
client = BatchClient(args.token_file, args.cache_dir)
|
|
outcome = client.run(request)
|
|
record = outcome["record"]
|
|
result = record["result"]
|
|
content = json.loads(result["response"])
|
|
if args.mode == "profile":
|
|
validate_profile(content, sources[0])
|
|
elif args.mode == "group":
|
|
validate_families(content, sources)
|
|
elif content != {"status": "LOCAL_OK"}:
|
|
raise ValueError("synthetic transport check failed")
|
|
print(json.dumps({"synthetic": True, "mode": args.mode, "model": args.model,
|
|
"cache_hit": outcome["cache_hit"], "result_file": outcome["path"],
|
|
"wall_seconds": record["client_wall_seconds"],
|
|
"load_seconds": result.get("load_duration", 0) / 1e9,
|
|
"prompt_tokens": result.get("prompt_eval_count"),
|
|
"output_tokens": result.get("eval_count"), "contract_valid": True}), flush=True)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
main()
|