1. 参数问题

File "/home/wuwenliang/py_workspace/llm_finetune/qwenvl_latext_ocr/train_latex_ocr_sft_yaml_3.py", line 260, in <module>
trainer.train()
File "/data2/llmtuner/lib/python3.10/site-packages/transformers/trainer.py", line 2238, in train
return inner_training_loop(
File "/data2/llmtuner/lib/python3.10/site-packages/transformers/trainer.py", line 2698, in _inner_training_loop
self._maybe_log_save_evaluate(
File "/data2/llmtuner/lib/python3.10/site-packages/transformers/trainer.py", line 3138, in _maybe_log_save_evaluate
is_new_best_metric = self._determine_best_metric(metrics=metrics, trial=trial)
File "/data2/llmtuner/lib/python3.10/site-packages/transformers/trainer.py", line 3208, in _determine_best_metric
raise KeyError(
KeyError("The 
metric_for_best_model
 training argument is set to 'eval_accuracy', which is not found in the evaluation metrics. The available evaluation metrics are: ['eval_loss', 'eval_runtime', 'eval_samples_per_second', 'eval_steps_per_second', 'epoch']. Consider changing the 
metric_for_best_model
 via the TrainingArguments.")
17%|█▋        | 62/372 [03:53<19:27,  3.77s/it]
Exception ignored in: <function tqdm.del at 0x7f91837d6f80>
Traceback (most recent call last):
File "/data2/llmtuner/lib/python3.10/site-packages/tqdm/std.py", line 1148, in del
File "/data2/llmtuner/lib/python3.10/site-packages/tqdm/std.py", line 1303, in close
File "/data2/llmtuner/lib/python3.10/site-packages/tqdm/std.py", line 1287, in fp_write
File "/data2/llmtuner/lib/python3.10/site-packages/tqdm/utils.py", line 196, in inner
File "/data2/llmtuner/lib/python3.10/site-packages/swanlab/log/log.py", line 93, in write_handler
File "/data2/llmtuner/lib/python3.10/site-packages/swankit/env.py", line 90, in create_time
AttributeError: 'NoneType' object has no attribute 'datetime'

2. 分析问题

这个错误表明你设置了 metric_for_best_model="eval_accuracy" ，但在评估过程中并没有计算 accuracy 指标。可用的评估指标只有： ['eval_loss', 'eval_runtime', 'eval_samples_per_second', 'eval_steps_per_second', 'epoch'] 。

3. 解决方案

方案1：使用存在的评估指标（推荐）

args = TrainingArguments(
    # ... 其他参数 ...
    metric_for_best_model="eval_loss",  # 使用存在的评估指标
    greater_is_better=False,            # 对于 loss，越小越好
    load_best_model_at_end=True,
)

方案2：添加 accuracy 计算函数

如果你确实需要计算 accuracy，需要在训练器中添加计算函数：


import numpy as np
from transformers import EvalPrediction

def compute_metrics(eval_pred):
    predictions, labels = eval_pred
    # 根据你的任务类型实现 accuracy 计算
    # 例如对于分类任务：
    predictions = np.argmax(predictions, axis=1)
    accuracy = (predictions == labels).astype(np.float32).mean()
    return {"accuracy": accuracy}

# 在创建 Trainer 时添加 compute_metrics
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    compute_metrics=compute_metrics,  # 添加这个
    # ... 其他参数 ...
)

方案3：检查当前代码中的 compute_metrics

查看你的代码中是否已经定义了 
compute_metrics
 函数，但可能实现有问题或者没有正确传递给 Trainer。


推荐修改

根据你的错误信息，最简单的解决方案是：


args = TrainingArguments(
    # ... 其他参数 ...
    eval_strategy="steps",
    save_strategy="steps",
    eval_steps=500,
    save_steps=500,
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",  # 改为使用 eval_loss
    greater_is_better=False,            # loss 越小越好
    # ... 其他参数 ...
)

如果你需要 accuracy 指标

如果你的任务是分类任务，可以这样实现：


def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=-1)
    accuracy = np.mean(predictions == labels)
    return {"accuracy": accuracy}

trainer = Trainer(
    # ... 其他参数 ...
    compute_metrics=compute_metrics,
)

选择方案1是最快速的解决方案，先让训练正常运行起来。