Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.
Python calibrationclassificationdecision-modelhuggingfacejevmodernbertmultilingualnlppythonpytorchroutingtyped-decisionszero-shot
Merge pull request #951 from sharath-sms/feat/parallel-option-layout
Nandakishorc06208860921
diff --cc laya/train.pyindex f5f14708,e8a210f4..759ca0a0--- a/laya/train.py+++ b/laya/train.py@@@ -37,12 -34,10 +37,13 @@@ from typing import Any, Callable, Dictimport torch-from .calibrate import fit_temperature_map+from .calibrate import MIN_TYPE_N, fit_temperature_mapfrom .common import (+ OPTION_LAYOUTS,+ QTYPE_NAMES,QTYPES,+ TEMP_MAX,+ TEMP_MIN,build_head,build_model,build_sequence,@@@ -548,16 -498,10 +567,16 @@@ def finetune(data: str, model_dir: stros.path.join(output_dir, "checkpoint_latest"))history = train_model(model, tok, train_items, config, dev, max_len, head_max_len,- on_epoch_end=checkpoint_latest)+ on_epoch_end=checkpoint_latest, parallel=parallel)- records = calibration_records(model, tok, calib_items, dev, max_len, head_max_len)- fitted = fit_temperature_map(calibration_records(model, tok, calib_items, dev, max_len, head_max_len,- parallel=parallel))++ records = calibration_records(model, tok, calib_items, dev, max_len, head_max_len, parallel=parallel)+ fitted = fit_temperature_map(records)+ calibration = calibration_report(records, fitted["temperature"])+ for name, entry in calibration.items():+ for issue in entry["issues"]:+ # Logged as a warning rather than left to the report alone: a checkpoint whose+ # confidences were never calibrated looks the same as one that was.+ warnings.warn("laya.train: %s calibration: %s" % (name, issue), RuntimeWarning, stacklevel=2)out_cfg = dict(cfg, max_len=max_len, head_max_len=head_max_len, fine_tuned=True,temperature=fitted["temperature"])# An inherited bucket map takes precedence at inference and would mask the new fit.