patch adaptation: model import

requirements.txt

@@ -23,4 +23,5 @@ mosestokenizer
 cached_property
 tqdm
 skops
-pandas
+pandas
+protobuf==3.20

wtpsplit/train/train_lora.py

@@ -10,16 +10,18 @@
 from glob import glob
 from typing import List, Optional, Union
 
+import adapters
 import datasets
 import numpy as np
 import torch
+from adapters import AdapterArguments
+from adapters.models import MODEL_MIXIN_MAPPING
+from adapters.models.bert.mixin_bert import BertModelAdaptersMixin
 from tokenizers import AddedToken
 from tqdm import tqdm
 from transformers import AutoTokenizer, HfArgumentParser, TrainingArguments, set_seed
 
-import adapters
 import wandb
-from adapters import AdapterArguments
 from wtpsplit.models import SubwordXLMConfig, SubwordXLMForTokenClassification
 from wtpsplit.train.adaptertrainer import AdapterTrainer
 from wtpsplit.train.evaluate import evaluate_sentence
@@ -32,6 +34,8 @@
 
 os.environ["TOKENIZERS_PARALLELISM"] = "false"
 
+MODEL_MIXIN_MAPPING["SubwordXLMRobertaModel"] = BertModelAdaptersMixin
+
 
 @dataclass
 class Args:
@@ -387,29 +391,6 @@ def maybe_pad(text):
     for lang in tqdm(data.keys(), desc="Language"):
         if lang in args.include_languages:
             for dataset_name in data[lang]["sentence"].keys():
-                if "corrupted-asr" in dataset_name and (
-                    "lyrics" not in dataset_name
-                    and "short" not in dataset_name
-                    and "code" not in dataset_name
-                    and "ted2020" not in dataset_name
-                    and "legal" not in dataset_name
-                ):
-                    print("SKIP: ", lang, dataset_name)
-                    continue
-                if "legal" in dataset_name and not ("laws" in dataset_name or "judgements" in dataset_name):
-                    print("SKIP: ", lang, dataset_name)
-                    continue
-                if "social-media" in dataset_name:
-                    continue
-                if "nllb" in dataset_name:
-                    continue
-                if lang == "en" and "legal-all-laws" in dataset_name:
-                    # not available.
-                    print("SKIP: ", lang, dataset_name)
-                    continue
-                if not any(x in dataset_name for x in ["ersatz", "opus", "ud"]):
-                    print("SKIP: ", lang, dataset_name)
-                    continue
                 print("RUNNING:", dataset_name, lang)
                 # do model stuff here; otherwise, head params would be overwritten every time
                 backbone = SubwordXLMForTokenClassification.from_pretrained(