# KeyError: 'loss' even though my dataset has labels

**URL:** </static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fkeyerror-loss-even-though-my-dataset-has-labels%2F11563%3E%5C
**Category:** Beginners\
**Created:** [November 10, 2021, 1:23pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fkeyerror-loss-even-though-my-dataset-has-labels%2F11563 "2021-11-10T13:23:32Z")\
**Posts on this page:** 2\
**Page:** 1

<div class="post-metadata">

**Author:** ![marcomatta](https://avatars.discourse-cdn.com/v4/letter/m/34f0e0/32.png) [@marcomatta](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fmarcomatta)%5C
**Post date:** [November 10, 2021, 1:23pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fkeyerror-loss-even-though-my-dataset-has-labels%2F11563%2F1 "2021-11-10T13:23:32Z")

</div>

Hi everyone! I’m trying to fine-tune on a NER task the `Musixmatch/umberto-commoncrawl-cased-v1` model, on the italian section of the wikiann dataset. The notebook I’m looking up to is this: [notebooks/token\_classification.ipynb at master · huggingface/notebooks · GitHub](https://github.com/huggingface/notebooks/blob/master/examples/token_classification.ipynb).  
Dataset’s initial structure is:

```auto
DatasetDict({
    validation: Dataset({
        features: ['tokens', 'ner_tags', 'langs', 'spans'],
        num_rows: 10000
    })
    test: Dataset({
        features: ['tokens', 'ner_tags', 'langs', 'spans'],
        num_rows: 10000
    })
    train: Dataset({
        features: ['tokens', 'ner_tags', 'langs', 'spans'],
        num_rows: 20000
    })
})

```

It has no labels but the `DataCollatorForTokenClassification` should help me out generating them.

```auto
from transformers import DataCollatorForTokenClassification
from datasets import load_metric

data_collator = DataCollatorForTokenClassification(tokenizer)
metric = load_metric("seqeval")

```

```auto
from transformers import AutoModel, TrainingArguments, Trainer

model = AutoModel.from_pretrained("Musixmatch/umberto-commoncrawl-cased-v1")
model_name = model_checkpoint.split("/")[-1]

training_args = TrainingArguments(

    f"{model_name}-finetuned-{task}", # output directory

    evaluation_strategy = "epoch",

    num_train_epochs=3, # total number of training epochs

    learning_rate=2e-5,

    per_device_train_batch_size=batch_size, # batch size per device during training

    per_device_eval_batch_size=batch_size, # batch size for evaluation

    warmup_steps=500, # number of warmup steps for learning rate scheduler

    weight_decay=0.01, # strength of weight decay

    logging_dir='./logs', # directory for storing logs

    logging_steps=10,

)

trainer = Trainer(

    model,

    training_args,

    train_dataset=tokenized_dataset["train"],

    eval_dataset=tokenized_dataset["validation"],

    data_collator=data_collator,

    tokenizer=tokenizer,

    compute_metrics=compute_metrics

)

```

The error it raises when I run `trainer.train()` is:

```auto
KeyError Traceback (most recent call last)
<ipython-input-16-3435b262f1ae> in <module>()
----> 1 trainer.train()

3 frames
/usr/local/lib/python3.7/dist-packages/transformers/file_utils.py in __getitem__ (self, k)
   2041 if isinstance(k, str):
   2042 inner_dict = {k: v for (k, v) in self.items()}
-> 2043 return inner_dict[k]
   2044 else:
   2045 return self.to_tuple()[k]

KeyError: 'loss'

```

How can I fix it? What am I doig wrong? Thanks for the help!

---

<div class="post-metadata">

**Author:** ![sgugger](https://sea2.discourse-cdn.com/hellohellohello/user_avatar/discuss.huggingface.co/sgugger/32/2291_2.png) [@sgugger](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fsgugger)%5C
**Post date:** [November 10, 2021, 2:06pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fkeyerror-loss-even-though-my-dataset-has-labels%2F11563%2F2 "2021-11-10T14:06:58Z")

</div>

> [@marcomatta](#):
>
> It has no labels but the `DataCollatorForTokenClassification` should help me out generating them.

No, you need to preprocess your dataset to generate them. The data collator is only there to pad those labels as well as the inputs. Have a look at one of the token classification [example script](https://github.com/huggingface/transformers/tree/master/examples/pytorch/token-classification) or [example notebook](https://github.com/huggingface/notebooks/blob/master/examples/token_classification.ipynb)