From 128ff308a51cdb205a51b290399d62fa5a9ca49d Mon Sep 17 00:00:00 2001 From: "dingbin.wang" Date: Thu, 15 Oct 2026 16:35:55 +0800 Subject: [PATCH] explicit model classes + exc tail to termination-log --- train_jev.py | 21 ++++++++++++++++++--- 1 file changed, 18 insertions(+), 3 deletions(-) diff --git a/train_jev.py b/train_jev.py index fc3bb55..d83af80 100644 --- a/train_jev.py +++ b/train_jev.py @@ -9,7 +9,12 @@ import json, os, argparse, math, random import torch import torch.nn.functional as F from torch.utils.data import Dataset, DataLoader -from transformers import AutoTokenizer, AutoModelForSequenceClassification, get_cosine_schedule_with_warmup +from transformers import AutoTokenizer, get_cosine_schedule_with_warmup +try: + from transformers import XLMRobertaTokenizerFast as XTok +except Exception: + from transformers import XLMRobertaTokenizer as XTok +from transformers import XLMRobertaForSequenceClassification def load_jsonl(p): return [json.loads(l) for l in open(p, encoding="utf-8")] @@ -76,8 +81,18 @@ def main(): dev = "cuda" if torch.cuda.is_available() else "cpu" print(f"[jev] device={dev} base={args.base}", flush=True) - tok = AutoTokenizer.from_pretrained(args.base) - model = AutoModelForSequenceClassification.from_pretrained(args.base, num_labels=1).to(dev) + try: + tok = XTok.from_pretrained(args.base) + model = XLMRobertaForSequenceClassification.from_pretrained(args.base, num_labels=1).to(dev) + except Exception: + import traceback + tail = traceback.format_exc()[-1200:] + try: + with open("/dev/termination-log", "a") as f: + f.write("\nEXC-TAIL:\n" + tail) + except Exception: + pass + raise train = [r for r in load_jsonl(os.path.join(args.data, "train.jsonl")) if r.get("hard_label") in (0, 1)] calib = load_jsonl(os.path.join(args.data, "calibration.jsonl"))