![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154131-6803c42b42edb.png\")
<\/p>\n\u786c\u4ef6\u65b9\u9762\u9700\u8981\u4e00\u4e2a\u8f83\u9ad8\u914d\u7f6e\u7684GPU，\u6211\u8fd9\u91cc\u4f7f\u7528\u7684\u662fA100，80G\u663e\u5b58<\/p>\n
（\u7531\u4e8e\u670d\u52a1\u5668\u8fde\u4e0d\u4e0ahuggingface，\u6211\u5c06\u6a21\u578b\u4e0b\u8f7d\u597d\u5e76\u4e0a\u4f20\u5230\u670d\u52a1\u5668\u8fdb\u884c\u672c\u5730\u8bad\u7ec3）<\/p>\n
\u6a21\u578b\u5730\u5740：https:\/\/huggingface.co\/mistralai\/Ministral-8B-Instruct-2410<\/p>\n
\u4e0b\u8f7d\u6a21\u578b\u4e4b\u524d\u9700\u8981\u5148\u7533\u8bf7\u4e00\u4e0b\u6743\u9650，\u901a\u8fc7\u4e4b\u540e，\u5728Files and versions\u4e2d\u4e0b\u8f7d\u6a21\u578b\u76f8\u5173\u6587\u4ef6，\u6211\u76f4\u63a5\u4e0b\u8f7d\u4e86\u6240\u6709\u7684\u6587\u4ef6\u3002<\/p>\n
<\/p>\n
\u4e0b\u8f7d\u597d\u6a21\u578b\u4e4b\u540e，\u53ef\u4ee5\u901a\u8fc7scp\u6307\u4ee4\u4e0a\u4f20\u5230\u670d\u52a1\u5668\u4e2d，\u6bd4\u5982\u6211\u4e0a\u4f20\u5230\u670d\u52a1\u5668\u4e0a\u7684\u4f4d\u7f6e\u4e3a：\/home\/user_m\/models\/Ministral-8B-Instruct-2410<\/p>\n
\u4e0e\u6a21\u578b\u4e00\u6837，\u7531\u4e8e\u65e0\u6cd5\u8fde\u63a5\u5230huggingface，\u6211\u9009\u62e9\u5c06\u6570\u636e\u96c6\u4e0b\u8f7d\u5e76\u4e0a\u4f20\u5230\u670d\u52a1\u5668\u4e2d，\u6211\u8fd9\u91cc\u4f7f\u7528\u7684\u6570\u636e\u96c6\u4e3a\u00a0e-SNLI: <\/span>https:\/\/<\/span>huggingface.co<\/span>\/datasets\/<\/span>esnli<\/span>\/<\/span>esnli<\/span><\/p>\n \u540c\u6837，\u5c06Files and versions\u4e2d\u7684\u6587\u4ef6\u4e0b\u8f7d\u5e76\u901a\u8fc7scp\u6307\u4ee4\u4e0a\u4f20\u5230\u670d\u52a1\u5668，\u6211\u7684\u6570\u636e\u96c6\u5728\u670d\u52a1\u5668\u4e2d\u7684\u4f4d\u7f6e\u4e3a：\/home\/user_m\/DataSet\/esnil<\/p>\n \u4e0b\u9762\u662f\u6570\u636e\u96c6\u4e2d\u4e00\u6761\u6570\u636e\u7684\u793a\u4f8b：<\/p>\n { \u6570\u636e\u96c6\u5185\u5bb9：<\/p>\n peft\u5730\u5740：https:\/\/huggingface.co\/docs\/peft\/index<\/p>\n \u5fae\u8c03\u4ee3\u7801\u5982\u4e0b： \u00a0<\/p>\n from copy import deepcopy POS_WEIGHT, NEG_WEIGHT, NATURAL_WEIGHT = (1.1637114032405993, 0.8766697374481806, 1.0)<\/p>\n def get_args(): arguments = parser.parse_args() def compute_metrics(eval_pred): logits, labels = eval_pred class CustomCallback(TrainerCallback): def on_epoch_end(self, args, state, control, **kwargs): def get_dataset_and_collator( Paramters: tokenizer = AutoTokenizer.from_pretrained( def _preprocesscing_function(examples): col_to_delete = ['explanation_1', 'explanation_2','explanation_3', 'premise','hypothesis'] padding_collator = DataCollatorWithPadding(tokenizer=tokenizer)<\/p>\n return tokenized_datasets, padding_collator<\/p>\n def get_lora_model(model_checkpoints, num_labels=3, rank=4, alpha=16, lora_dropout=0.1, bias='none'): peft_config = LoraConfig( model = get_peft_model(model, peft_config)<\/p>\n print(model.print_trainable_parameters())<\/p>\n return model<\/p>\n def get_weighted_trainer(pos_weight, neg_weight,neutral_weight=1.0): def main(args): dataset, collator = get_dataset_and_collator( training_args = TrainingArguments( model = get_lora_model( # move model to GPU device weighted_trainer = get_weighted_trainer(POS_WEIGHT, NEG_WEIGHT,NATURAL_WEIGHT)<\/p>\n trainer = weighted_trainer( if __name__ == "__main__": \u539f\u56e0：<\/p>\n \u5728\u52a0\u8f7dtokenizer\u65f6，\u4f7f\u7528\u4e86padding<\/p>\n tokenizer.pad_token = tokenizer.eos_token \u867d\u7136\u6307\u5b9a\u4e86tokenizer\u7684pad_token,\u4f46\u662f\u8fd9\u4e2a\u6a21\u578b\u672c\u8eab\u5e76\u6ca1\u6709\u9ed8\u8ba4\u7684pad_token_id，\u5bfc\u81f4\u6a21\u578b\u8ba4\u4e0d\u51fa\u8fd9\u4e2apad_token\u6240\u4ee5\u51fa\u73b0\u4e86\u62a5\u9519.<\/p>\n<\/p>\n \u89e3\u51b3\u529e\u6cd5：<\/p>\n \u5728\u52a0\u8f7d\u5b8c\u6a21\u578b\u4e4b\u540e，\u6307\u5b9apad_token_id:<\/p>\n model = AutoModelForSequenceClassification.from_pretrained( \u8fd9\u6837\u6a21\u578b\u5c31\u77e5\u9053\u4f7f\u7528\u4e86eos_token\u4f5c\u4e3apad_token<\/p>\n<\/p>\n \u8bc4\u4f30\u5fae\u8c03\u4e4b\u540e\u6a21\u578b\u63a8\u7406\u7684accuracy\u65f6，\u53d1\u73b0\u6b63\u786e\u7387\u53ea\u67090.3\u591a\u4e00\u70b9，\u5bf9\u4e8e\u4e09\u5206\u7c7b\u95ee\u9898\u6765\u8bf4，\u8fd9\u4e2a\u51c6\u786e\u7387\u548c\u778e\u731c\u6ca1\u6709\u4ec0\u4e48\u533a\u522b\u3002\u3002\u3002\u6211\u5206\u6790\u4e86\u6211\u6700\u5f00\u59cb\u8fdb\u884c\u8bad\u7ec3\u7684\u4ee3\u7801，\u53d1\u73b0\u4e86\u5982\u4e0b\u95ee\u9898：<\/p>\n \u6700\u521d\u6211\u8fdb\u884c\u8bad\u7ec3\u7684\u4ee3\u7801\u5982\u4e0b： \u00a0<\/p>\n import torch from peft import get_peft_model, LoraConfig import os<\/p>\n model_path = "\/home\/user_m\/models\/Ministral-8B-Instruct-2410" print(tokenizer.pad_token_id) def preprocess_function(examples): # \u4f7f\u7528Peft\u5fae\u8c03\u6a21\u578b # \u83b7\u53d6 PEFT \u6a21\u578b # \u8bbe\u7f6e\u8bad\u7ec3\u53c2\u6570 # \u4f7f\u7528 Trainer \u8fdb\u884c\u5fae\u8c03 # \u5f00\u59cb\u8bad\u7ec3 # \u4f7f\u7528 Trainer \u8fdb\u884c\u8bc4\u4f30 \u7136\u540e\u6bcf\u6b21\u52a0\u8f7d\u5b8c\u6a21\u578b\u4e4b\u540e\u4f1a\u51fa\u73b0\u4ee5\u4e0b\u4fe1\u606f：<\/p>\n Some weights of the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 were not used when initializing MistralForSequenceClassification: ['lm_head.weight']<\/p>\n Some weights of MistralForSequenceClassification were not initialized from the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 and are newly initialized: ['score.weight'] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference. \u95ee\u9898\u89e3\u51b3：<\/p>\n \u9996\u5148\u5728\u4ee3\u7801\u5f00\u5934\u52a0\u5165\u5982\u4e0b\u5185\u5bb9：<\/p>\n from transformers import logging<\/p>\n logging.set_verbosity_info() <\/p>\n \u8fd9\u884c\u6307\u4ee4\u53ef\u4ee5\u4f7f\u5f97\u7ec8\u7aef\u8f93\u51fa\u7684\u4fe1\u606f\u66f4\u4e3a\u8be6\u7ec6，\u518d\u6b21\u8fd0\u884c\u4ee3\u7801，\u52a0\u8f7d\u5b8c\u6a21\u578b\u4e4b\u540e\u539f\u672c\u7684\u4fe1\u606f\u53d8\u4e3a： \u5927\u81f4\u610f\u601d\u662f\u8bf4，\u52a0\u8f7d\u6a21\u578b\u65f6，\u6ca1\u6709\u4f7f\u7528lm_head_weight\u8fd9\u4e2a\u6743\u91cd，\u8fd9\u662f\u56e0\u4e3a\u6211\u4eec\u4f7f\u7528\u6a21\u578b\u8fdb\u884c\u4e86\u5176\u4ed6\u7684\u4efb\u52a1\u3002\u5e76\u4e14\u52a0\u8f7d\u6a21\u578b\u65f6score.weight\u8fd9\u4e2a\u6743\u91cd\u6ca1\u6709\u4ece\u6a21\u578b\u7684checkpoint\u4e2d\u627e\u5230，\u6240\u4ee5\u6a21\u578b\u5bf9\u8fd9\u4e00\u90e8\u5206\u7684\u6743\u91cd\u91cd\u65b0\u8fdb\u884c\u4e86\u968f\u673a\u521d\u59cb\u5316\u3002<\/p>\n \u6211\u4eec\u52a0\u8f7d\u5b8c\u6a21\u578b\u4e4b\u540e\u67e5\u770b\u6a21\u578b\u7684config：<\/p>\n print("model.config: ",model.config) <\/p>\n \u00a0\u53ef\u4ee5\u5f97\u5230\u5982\u4e0b\u8f93\u51fa\u4fe1\u606f：<\/p>\n Model config MistralConfig { \u91cd\u70b9\u5173\u6ce8"architectures": ["MistralForCausalLM"]，\u8bf4\u660e\u6a21\u578b\u672c\u8eab\u662f\u7528\u4e8e\u8fdb\u884cLM\u4efb\u52a1\u7684\u3002<\/p>\n \u6211\u4eec\u518d\u4f7f\u7528AutoModelForCausalLM\u52a0\u8f7d\u5e76\u67e5\u770b\u6a21\u578b：<\/p>\n model = AutoModelForCausalLM.from_pretrained(model_name) \u8fd9\u6b21\u6211\u4eec\u4e0d\u518d\u5f97\u5230Some weights of the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 were not used when initializing MistralForSequenceClassification: ['lm_head.weight']\u8fd9\u6837\u7684\u4fe1\u606f\u3002<\/p>\n \u5e76\u4e14\u4f1a\u5f97\u5230\u4ee5\u4e0b\u8f93\u51fa\u4fe1\u606f： \u53ef\u4ee5\u770b\u5230\u6a21\u578b\u7684\u6700\u540e\u4e00\u5c42\u662f\u4e00\u4e2alm_head\u7ebf\u6027\u5c42,\u8fd9\u4e00\u5c42\u5176\u5b9e\u5c31\u662f\u5bf9\u5e94\u6a21\u578b\u8fdb\u884cLM\u4efb\u52a1\u7528\u7684\u5934，\u5f53\u6211\u4eec\u4f7f\u7528AutoModelForCausalLM\u52a0\u8f7d\u6a21\u578b\u65f6，\u4e0e\u539f\u672c\u6a21\u578b\u4efb\u52a1\u4e00\u81f4，\u4e0d\u4f1a\u51fa\u73b0\u95ee\u9898\u3002<\/p>\n \u4f46\u662f\u5f53\u6211\u4eec\u4f7f\u7528AutoModelForSequenceClassification\u52a0\u8f7d\u6a21\u578b，\u5e76\u8f93\u51fa\u6a21\u578b\u7ed3\u6784\u4fe1\u606f\u65f6：<\/p>\n model = AutoModelForSequenceClassification.from_pretrained( \u5c31\u4f1a\u51fa\u73b0Some weights of the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 were not used when initializing MistralForSequenceClassification: ['lm_head.weight']\u4fe1\u606f，\u4e14\u6a21\u578b\u7ed3\u6784\u5982\u4e0b： \u6ce8\u610f，\u53ef\u4ee5\u770b\u5230\u6a21\u578b\u6700\u540e\u4e00\u5c42\u7531lm_head\u53d8\u6210\u4e86score，\u8fd9\u662f\u56e0\u4e3a\u4f7f\u7528AutoModelForSequenceClassification\u65f6，\u4f1a\u81ea\u52a8\u7ed9\u6a21\u578b\u9002\u914dscore\u5206\u7c7b\u5934\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1，\u6240\u4ee5\u518d\u6765\u5206\u6790\u63d0\u793a\u4fe1\u606f：<\/p>\n Some weights of the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 were not used when initializing MistralForSequenceClassification: ['lm_head.weight']<\/p>\n Some weights of MistralForSequenceClassification were not initialized from the model checkpoint at \/home\/user_m\/models\/Ministral-8B-Instruct-2410 and are newly initialized: ['score.weight'] You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.<\/p>\n \u7531\u4e8e\u6211\u4eec\u5c06\u6a21\u578b\u7528\u4e8eSequenceClassification\u4efb\u52a1，\u539f\u672c\u7684lm_head\u5c42\u88ab\u66ff\u6362\u6210\u4e86score\u5c42，\u8fd9\u6837\u4f7f\u5f97\u539f\u672ccheckpoint\u4e2dlm_head\u76f8\u5173\u6743\u91cd\u6ca1\u6cd5\u52a0\u8f7d，\u5e76\u4e14checkpoint\u4e2d\u539f\u672c\u5c31\u5e76\u4e0d\u5b58\u5728score\u6743\u91cd，\u6240\u4ee5\u624d\u4f1a\u51fa\u73b0\u4e0a\u8ff0\u4fe1\u606f\u63d0\u793a\u3002\u7531\u4e8e\u4ece\u65b0\u521d\u59cb\u5316\u4e86score\u5c42，\u6240\u4ee5\u5f53\u6a21\u578b\u8ba1\u7b97\u5230score\u5c42\u65f6，\u5229\u7528\u8fd9\u4e2a\u672a\u7ecf\u8bad\u7ec3\u7684\u5206\u7c7b\u5934\u505a\u6700\u540e\u4e00\u6b65\u8f93\u51fa，\u5f97\u5230\u7684\u7ed3\u679c\u7c7b\u4f3c\u4e8e\u968f\u673a\u9009\u62e9\u4e5f\u5e76\u4e0d\u5947\u602a\u3002\u3002\u3002<\/p>\n \u95ee\u9898\u89e3\u51b3：<\/p>\n \u5728\u5206\u6790\u4e86\u4ee3\u7801\u540e，\u6211\u8ba4\u4e3a\u95ee\u9898\u51fa\u5728peft_config\u8fd9\u91cc，\u8fd9\u91cc\u9700\u8981\u6307\u5b9a\u4efb\u52a1\u7c7b\u578btask_type=TaskType.SEQ_CLS<\/p>\n peft_config = LoraConfig( \u8fd9\u6837\u8bad\u7ec3\u4e4b\u540e，\u518d\u6b21\u4f7f\u7528AutoModelForSequenceClassification\u52a0\u8f7d\u6a21\u578b\u65f6\u867d\u7136\u8fd8\u662f\u4f1a\u51fa\u73b0\u63d0\u793a\u4fe1\u606f，\u4f46\u662f\u8fd9\u4e2a\u4fe1\u606f\u662f\u7531\u4e8e\u52a0\u8f7dbaseline\u6a21\u578b\u51fa\u73b0\u7684，\u5728\u52a0\u8f7d\u5b8cbaseline\u4e4b\u540e，\u4f1a\u518d\u52a0\u8f7dpeft\u76f8\u5173\u6743\u91cd\u4fe1\u606f，\u800cscore\u6743\u91cd\u4fe1\u606f\u4f1a\u7531peft\u8fdb\u884c\u4fdd\u5b58，\u6240\u4ee5\u8fdb\u884c\u8bc4\u4f30\u65f6： \u53d1\u73b0accuracy\u8fbe\u52300.86，\u662f\u4e00\u4e2a\u5408\u7406\u7684\u503c，\u8bf4\u660e\u5fae\u8c03\u6210\u529f，\u95ee\u9898\u89e3\u51b3！<\/p>\n \u603b\u7ed3\u6765\u8bf4，\u6a21\u578b\u5176\u5b9e\u662f\u5305\u542b\u4e24\u4e2a\u90e8\u5206，\u4e3b\u4f53\u6a21\u578b+\u5904\u7406\u76f8\u5173\u4efb\u52a1\u7684\u4e0b\u6e38\u6a21\u578b（\u6bd4\u5982lm_head\u3001score\u7b49），\u5c06\u672c\u6765\u7528\u4e8e\u4efb\u52a1A\u7684\u6a21\u578b\u5e94\u7528\u4e8e\u4efb\u52a1B\u65f6，\u5176\u5b9e\u662f\u4f1a\u5c06\u5904\u7406A\u7684\u4e0b\u6e38\u6a21\u578b\u66ff\u6362\u4e3a\u5904\u7406B\u7684\u4e0b\u6e38\u6a21\u578b，\u6240\u4ee5\u5982\u679c\u6211\u4eec\u5e0c\u671b\u6a21\u578b\u80fd\u6709\u826f\u597d\u7684\u8868\u73b0，\u5c31\u9700\u8981\u786e\u4fdd\u66ff\u6362\u540e\u7684\u4e0b\u6e38\u6a21\u578b\u4e5f\u662f\u7ecf\u8fc7\u826f\u597d\u8bad\u7ec3\u800c\u4e0d\u662f\u968f\u673a\u521d\u59cb\u5316\u7684，\u8fd9\u4e00\u6b65\u9700\u8981\u5728\u5fae\u8c03\u6a21\u578b\u6216\u8bad\u7ec3\u6a21\u578b\u65f6\u683c\u5916\u6ce8\u610f\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb899\u6b21\uff0c\u70b9\u8d5e20\u6b21\uff0c\u6536\u85cf30\u6b21\u3002\u8fd9\u662f\u56e0\u4e3a\u4f7f\u7528AutoModelForSequenceClassification\u65f6\uff0c\u4f1a\u81ea\u52a8\u7ed9\u6a21\u578b\u9002\u914dscore\u5206\u7c7b\u5934\u7528\u4e8e\u5206\u7c7b\u4efb\u52a1\uff0c\u6240\u4ee5\u518d\u6765\u5206\u6790\u63d0\u793a\u4fe1\u606f\uff1a_ministral 8b<\/p>\n","protected":false},"author":2,"featured_media":28305,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,86],"topic":[],"class_list":["post-28313","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-server","tag-50","tag-86"],"yoast_head":"\n
\n "explanation_1": "A woman must be present to smile.",
\n "explanation_2": "A woman smiling implies that she is present.",
\n "explanation_3": "A smiling woman is also present.",
\n "hypothesis": "A woman is present.",
\n "label": 0,
\n "premise": "A woman smiles at the child."
\n} <\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154132-6803c42c8cff6.png\")
<\/li>\n\n
2. \u4f7f\u7528peft\u8fdb\u884c\u5fae\u8c03<\/h3>\n
\nfrom argparse import ArgumentParser
\nfrom datasets import load_from_disk,load_dataset
\nimport evaluate
\nimport numpy as np
\nfrom peft import get_peft_model, LoraConfig, TaskType
\nfrom transformers import AutoTokenizer, DataCollatorWithPadding
\nfrom transformers import AutoModelForSequenceClassification
\nfrom transformers import TrainingArguments, Trainer, TrainerCallback
\nimport torch<\/p>\n
\n parser = ArgumentParser(description="Fine-tune an LLM model with PEFT")
\n parser.add_argument(
\n "–data_path",
\n type=str,
\n default='\/home\/user_m\/DataSet\/esnil\/esnli.py',
\n required=True,
\n help="Path to Huggingface pre-processed dataset",
\n )
\n parser.add_argument(
\n "–output_path",
\n type=str,
\n default='\/home\/user_m\/output',
\n required=True,
\n help="Path to store the fine-tuned model",
\n )
\n parser.add_argument(
\n "–model_name",
\n type=str,
\n default='\/home\/user_m\/models\/Ministral-8B-Instruct-2410',
\n required=True,
\n help="Name of the pre-trained LLM to fine-tune",
\n )
\n parser.add_argument(
\n "–max_length",
\n type=int,
\n default=128,
\n required=False,
\n help="Maximum length of the input sequences",
\n )
\n parser.add_argument(
\n "–set_pad_id",
\n action="store_true",
\n help="Set the id for the padding token, needed by models such as Mistral-7B",
\n )
\n parser.add_argument(
\n "–lr", type=float, default=2e-4, help="Learning rate for training"
\n )
\n parser.add_argument(
\n "–train_batch_size", type=int, default=64, help="Train batch size"
\n )
\n parser.add_argument(
\n "–eval_batch_size", type=int, default=64, help="Eval batch size"
\n )
\n parser.add_argument(
\n "–num_epochs", type=int, default=2, help="Number of epochs"
\n )
\n parser.add_argument(
\n "–weight_decay", type=float, default=0.1, help="Weight decay"
\n )
\n parser.add_argument(
\n "–lora_rank", type=int, default=4, help="Lora rank"
\n )
\n parser.add_argument(
\n "–lora_alpha", type=float, default=0.0, help="Lora alpha"
\n )
\n parser.add_argument(
\n "–lora_dropout", type=float, default=0.2, help="Lora dropout"
\n )
\n parser.add_argument(
\n "–lora_bias",
\n type=str,
\n default='none',
\n choices={"lora_only", "none", 'all'},
\n help="Layers to add learnable bias"
\n )<\/p>\n
\n return arguments<\/p>\n
\n precision_metric = evaluate.load("\/home\/user_m\/evaluate\/metrics\/precision")
\n recall_metric = evaluate.load("\/home\/user_m\/evaluate\/metrics\/recall")
\n f1_metric = evaluate.load("\/home\/user_m\/evaluate\/metrics\/f1")
\n accuracy_metric = evaluate.load("\/home\/user_m\/evaluate\/metrics\/accuracy")<\/p>\n
\n predictions = np.argmax(logits, axis=-1)
\n precision = precision_metric.compute(predictions=predictions, references=labels,average='macro')["precision"]
\n recall = recall_metric.compute(predictions=predictions, references=labels,average='macro')["recall"]
\n f1 = f1_metric.compute(predictions=predictions, references=labels,average='macro')["f1"]
\n accuracy = accuracy_metric.compute(predictions=predictions, references=labels)["accuracy"]
\n return {"precision": precision, "recall": recall, "f1-score": f1, 'accuracy': accuracy}<\/p>\n
\n def __init__(self, trainer) -> None:
\n super().__init__()
\n self._trainer = trainer<\/p>\n
\n if control.should_evaluate:
\n control_copy = deepcopy(control)
\n self._trainer.evaluate(eval_dataset=self._trainer.train_dataset, metric_key_prefix="train")
\n return control_copy<\/p>\n
\n data_path,
\n model_checkpoints,
\n # add_prefix_space=True,
\n max_length=128,
\n truncation=True,
\n set_pad_id=False
\n):
\n """
\n Load the preprocessed HF dataset with train, valid and test objects<\/p>\n
\n ———
\n data_path: str
\n Path to the pre-processed HuggingFace dataset
\n model_checkpoints:
\n Name of the pre-trained model to use for tokenization
\n """
\n data = load_dataset(data_path)<\/p>\n
\n model_checkpoints,
\n # add_prefix_space=add_prefix_space
\n )
\n print("Tokenizer loaded successfully")
\n tokenizer.pad_token = tokenizer.eos_token<\/p>\n
\n return tokenizer(examples["premise"],examples["hypothesis"], truncation=True, padding="max_length", max_length=max_length)<\/p>\n
\n tokenized_datasets = data.map(_preprocesscing_function, batched=False)
\n tokenized_datasets = tokenized_datasets.remove_columns(col_to_delete)
\n tokenized_datasets.set_format("torch")<\/p>\n
\n """
\n TODO
\n """
\n model = AutoModelForSequenceClassification.from_pretrained(
\n pretrained_model_name_or_path=model_checkpoints,
\n num_labels=num_labels,
\n device_map="auto",
\n offload_folder="offload",
\n trust_remote_code=True,
\n )
\n model.config.pad_token_id = model.config.eos_token_id<\/p>\n
\n task_type=TaskType.SEQ_CLS, r=rank, lora_alpha=alpha, lora_dropout=lora_dropout, bias=bias,
\n target_modules=[
\n "q_proj",
\n "v_proj",
\n # "score"
\n ],
\n )<\/p>\n
\n class _WeightedBCELossTrainer(Trainer):
\n def compute_loss(self, model, inputs, return_outputs=False,num_items_in_batch=None):
\n labels = inputs.pop("labels")
\n # forward pass
\n outputs = model(**inputs)
\n logits = outputs.get("logits")
\n # compute custom loss (suppose one has 3 labels with different weights)
\n loss_fct = torch.nn.CrossEntropyLoss(weight=torch.tensor([neg_weight, pos_weight, neutral_weight], device=labels.device, dtype=logits.dtype))
\n loss = loss_fct(logits.view(-1, self.model.config.num_labels), labels.view(-1))
\n return (loss, outputs) if return_outputs else loss
\n return _WeightedBCELossTrainer<\/p>\n
\n """
\n Training function
\n """<\/p>\n
\n args.data_path,
\n args.model_name,
\n max_length=args.max_length,
\n set_pad_id=args.set_pad_id,
\n # add_prefix_space=True,
\n truncation=True,
\n )<\/p>\n
\n output_dir=args.output_path,
\n learning_rate=args.lr,
\n lr_scheduler_type= "cosine",
\n warmup_ratio= 0.1,
\n per_device_train_batch_size=args.train_batch_size,
\n per_device_eval_batch_size=args.eval_batch_size,
\n num_train_epochs=args.num_epochs,
\n weight_decay=args.weight_decay,
\n evaluation_strategy="epoch",
\n save_strategy="epoch",
\n load_best_model_at_end=True,
\n gradient_checkpointing=True,
\n fp16=True,
\n max_grad_norm= 0.3,
\n )<\/p>\n
\n args.model_name,
\n rank=args.lora_rank,
\n alpha=args.lora_alpha,
\n lora_dropout=args.lora_dropout,
\n bias=args.lora_bias
\n )
\n if args.set_pad_id:
\n model.config.pad_token_id = model.config.eos_token_id<\/p>\n
\n if model.device.type != 'cuda':
\n model=model.to('cuda')<\/p>\n
\n model=model,
\n args=training_args,
\n train_dataset=dataset['train'],
\n eval_dataset=dataset["validation"],
\n data_collator=collator,
\n compute_metrics=compute_metrics
\n )
\n trainer.add_callback(CustomCallback(trainer))
\n trainer.train()<\/p>\n
\n args = get_args()
\n main(args) <\/p>\n3.\u6574\u4e2a\u8fc7\u7a0b\u4e2d\u9047\u5230\u7684\u95ee\u9898<\/h3>\n
3.1\u00a0 raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") ValueError: Cannot handle batch sizes > 1 if no padding token is defined.<\/h4>\n
![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154132-6803c42cbfda4.png\")
<\/p>\n
\ntokenizer(examples["premise"],examples["hypothesis"], truncation=True, padding="max_length", max_length=max_length) <\/p>\n
\n pretrained_model_name_or_path=model_checkpoints,
\n num_labels=num_labels,
\n device_map="auto",
\n offload_folder="offload",
\n trust_remote_code=True,
\n )
\nmodel.config.pad_token_id = model.config.eos_token_id <\/p>\n3.2 \u8bad\u7ec3\u540e\u8bc4\u4f30accuracy\u5f88\u4f4e<\/h4>\n
\nfrom transformers import AutoTokenizer, AutoModelForSequenceClassification
\nfrom datasets import load_dataset<\/p>\n
\nfrom transformers import Trainer, TrainingArguments<\/p>\n
\ntokenizer = AutoTokenizer.from_pretrained(model_path)<\/p>\n
\nif tokenizer.pad_token_id is None:
\n tokenizer.pad_token_id = tokenizer.eos_token_id
\n# \u52a0\u8f7d ESNLI \u6570\u636e\u96c6
\ndataset = load_dataset("\/home\/user_m\/DataSet\/esnil\/esnli.py",trust_remote_code=True)<\/p>\n
\n # print(examples["label"]) # \u6253\u5370\u6807\u7b7e\u67e5\u770b
\n return tokenizer(examples["premise"], examples["hypothesis"], padding='max_length' ,max_length = 128, truncation=True)
\n # return tokenizer(examples["premise"], examples["hypothesis"])
\nencoded_dataset = dataset.map(preprocess_function, batched=True)
\nprint(encoded_dataset)<\/p>\n
\n# \u8bbe\u7f6e LoRA \u914d\u7f6e
\nmodel = AutoModelForSequenceClassification.from_pretrained(model_path, num_labels=3) # ESNLI \u662f\u4e09\u5206\u7c7b\u95ee\u9898
\nmodel.config.pad_token_id = model.config.eos_token_id
\nlora_config = LoraConfig(
\n r=8, # LoRA\u7684\u4f4e\u79e9\u5927\u5c0f
\n lora_alpha=16, # \u6807\u91cf
\n lora_dropout=0.1,
\n target_modules=["q_proj", "k_proj", "v_proj"], # LoRA\u9002\u7528\u4e8e\u7684\u6a21\u5757
\n)<\/p>\n
\npeft_model = get_peft_model(model, lora_config)
\npeft_model.print_trainable_parameters()<\/p>\n
\ntraining_args = TrainingArguments(
\n output_dir="\/home\/user_m\/output\/round2",
\n evaluation_strategy="epoch",
\n save_strategy="epoch",
\n learning_rate=1e-6,
\n per_device_train_batch_size=32,
\n per_device_eval_batch_size=32,
\n fp16=True,
\n num_train_epochs=1,
\n weight_decay=0.01,
\n logging_dir='\/home\/user_m\/logs',
\n logging_steps=10,
\n)<\/p>\n
\ntrainer = Trainer(
\n model=peft_model,
\n args=training_args,
\n train_dataset=encoded_dataset["train"],
\n eval_dataset=encoded_dataset["validation"],
\n tokenizer = tokenizer,
\n)<\/p>\n
\ntrainer.train()<\/p>\n
\nresults = trainer.evaluate()<\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154132-6803c42ce92c9.png\")
<\/p>\n<\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154133-6803c42d2d9a5.png\")
<\/p>\n
\n "_name_or_path": "\/home\/user_m\/models\/Ministral-8B-Instruct-2410",
\n "architectures": [
\n "MistralForCausalLM"
\n ],
\n "attention_dropout": 0.0,
\n "bos_token_id": 1,
\n "eos_token_id": 2,
\n "head_dim": 128,
\n "hidden_act": "silu",
\n "hidden_size": 4096,
\n "id2label": {
\n "0": "LABEL_0",
\n "1": "LABEL_1",
\n "2": "LABEL_2"
\n },
\n "initializer_range": 0.02,
\n "intermediate_size": 12288,
\n "label2id": {
\n "LABEL_0": 0,
\n "LABEL_1": 1,
\n "LABEL_2": 2
\n },
\n "max_position_embeddings": 32768,
\n "model_type": "mistral",
\n "num_attention_heads": 32,
\n "num_hidden_layers": 36,
\n "num_key_value_heads": 8,
\n "rms_norm_eps": 1e-05,
\n "rope_theta": 100000000.0,
\n "sliding_window": 32768,
\n "tie_word_embeddings": false,
\n "torch_dtype": "bfloat16",
\n "transformers_version": "4.46.3",
\n "use_cache": true,
\n "vocab_size": 131072
\n} <\/p>\n
\nprint(model) <\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154133-6803c42d8c84c.png\")
<\/p>\n
\n pretrained_model_name_or_path=model_checkpoints,
\n num_labels=num_labels,
\n device_map="auto",
\n offload_folder="offload",
\n trust_remote_code=True,
\n )
\nprint(model) <\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154133-6803c42ddc6dd.png\")
<\/p>\n
\n task_type=TaskType.SEQ_CLS, r=rank, lora_alpha=alpha, lora_dropout=lora_dropout, bias=bias,
\n target_modules=[
\n "q_proj",
\n "v_proj",
\n # "score"
\n ],
\n ) <\/p>\n![\"\"](\"https:\/\/www.wsisp.com\/helps\/wp-content\/uploads\/2025\/04\/20250419154134-6803c42e5aa6e.png\")
<\/p>\n