-
Notifications
You must be signed in to change notification settings - Fork 120
models distilbert base cased
DistilBERT, a transformers model, is designed to be smaller and quicker than BERT. It underwent pretraining on the same dataset in a self-supervised manner, utilizing the BERT base model as a reference. This entails training solely on raw texts, without human annotation, thus enabling the utilization of vast amounts of publicly accessible data. An automated process generates inputs and labels from these texts, guided by the BERT base model. Specifically, the pretraining process involved three objectives:
Distillation loss: The model was trained to produce probabilities akin to those of the BERT base model. Masked language modeling (MLM): This constitutes a segment of the original training loss in the BERT base model. By randomly masking 15% of the words in a sentence, the model processes the entire masked sentence and endeavors to predict the masked words. This methodology differs from traditional recurrent neural networks (RNNs) or autoregressive models like GPT, which handle words sequentially or internally mask future tokens. MLM facilitates the acquisition of a bidirectional sentence representation by the model. Cosine embedding loss: The model was also trained to generate hidden states that closely resemble those of the BERT base model. In this manner, the model acquires a comparable internal representation of the English language to that of its teacher model, while being more efficient for inference or subsequent tasks.
DistilBERT was pretrained on the same data as BERT, which includes the BookCorpus dataset (consisting of 11,038 unpublished books) and English Wikipedia (excluding lists, tables, and headers).
The texts are lowercased and tokenized using WordPiece with a vocabulary size of 30,000. The model inputs are structured as follows: [CLS] Sentence A [SEP] Sentence B [SEP] With a 50% probability, Sentence A and Sentence B correspond to two consecutive sentences from the original corpus. Otherwise, a random sentence from the corpus is used. The combined length of the two “sentences” must be less than 512 tokens. Masking procedure for each sentence: 15% of tokens are masked. In 80% of cases, masked tokens are replaced by [MASK]. In 10% of cases, masked tokens are replaced by a different random token. In the remaining 10%, masked tokens remain unchanged.
| Task | Use case | Dataset | Python sample (Notebook) | CLI with YAML |
|---|---|---|---|---|
| Fill Mask | Fill Mask | rcds/wikipedia-for-mask-filling | evaluate-model-fill-mask.ipynb | evaluate-model-fill-mask.yml |
The model was trained on 8 NVIDIA V100 GPUs (each with 16 GB memory) for 90 hours. Refer to the training code for detailed hyperparameters.
When fine-tuned on downstream tasks, this model achieves the following results:
Glue test results:
| Task | MNLI | QQP | QNLI | SST-2 | CoLA | STS-B | MRPC | RTE |
|---|---|---|---|---|---|---|---|---|
| 82.2 | 88.5 | 89.2 | 91.3 | 51.3 | 85.8 | 87.5 | 59.9 |
While the training data for this model is generally neutral, it can still produce biased predictions. Additionally, it inherits some of the biases from its teacher model.
| Evaluation type | Python sample |
|---|---|
| Real time | sdk-example.ipynb](https://github.com/Azure/azureml-examples/blob/main/sdk/python/foundation-models/system/evaluation/fill-mask/fill-mask.ipynb) |
| Inference type | Python sample |
|---|---|
| Real time | sdk-example.ipynb |
| Real time | fill-mask-online-endpoint.ipynb |
{
"input_data": ["Paris is [MASK] of France"]
} ["part"]Version: 13
license : apache-2.0 task : fill-mask SharedComputeCapacityEnabled training_datasets : bookcorpus, wikipedia huggingface_model_id : distilbert-base-cased hiddenlayerscanned evaluation_compute_allow_list : ['Standard_DS4_v2', 'Standard_D8a_v4', 'Standard_D8as_v4', 'Standard_DS5_v2', 'Standard_DS12_v2', 'Standard_D16a_v4', 'Standard_D16as_v4', 'Standard_D32a_v4', 'Standard_D32as_v4', 'Standard_D48a_v4', 'Standard_D48as_v4', 'Standard_D64a_v4', 'Standard_D64as_v4', 'Standard_D96a_v4', 'Standard_D96as_v4', 'Standard_FX4mds', 'Standard_FX12mds', 'Standard_F16s_v2', 'Standard_F32s_v2', 'Standard_F48s_v2', 'Standard_F64s_v2', 'Standard_F72s_v2', 'Standard_FX24mds', 'Standard_FX36mds', 'Standard_FX48mds', 'Standard_E4s_v3', 'Standard_E8s_v3', 'Standard_E16s_v3', 'Standard_E32s_v3', 'Standard_E48s_v3', 'Standard_E64s_v3', 'Standard_NC4as_T4_v3', 'Standard_NC6s_v3', 'Standard_NC8as_T4_v3', 'Standard_NC12s_v3', 'Standard_NC16as_T4_v3', 'Standard_NC24s_v3', 'Standard_NC64as_T4_v3', 'Standard_NC24ads_A100_v4', 'Standard_NC48ads_A100_v4', 'Standard_NC96ads_A100_v4', 'Standard_ND96asr_v4', 'Standard_ND96amsr_A100_v4', 'Standard_ND40rs_v2'] inference_compute_allow_list : ['Standard_DS3_v2', 'Standard_D4a_v4', 'Standard_D4as_v4', 'Standard_DS4_v2', 'Standard_D8a_v4', 'Standard_D8as_v4', 'Standard_DS5_v2', 'Standard_D16a_v4', 'Standard_D16as_v4', 'Standard_D32a_v4', 'Standard_D32as_v4', 'Standard_D48a_v4', 'Standard_D48as_v4', 'Standard_D64a_v4', 'Standard_D64as_v4', 'Standard_D96a_v4', 'Standard_D96as_v4', 'Standard_F4s_v2', 'Standard_FX4mds', 'Standard_F8s_v2', 'Standard_FX12mds', 'Standard_F16s_v2', 'Standard_F32s_v2', 'Standard_F48s_v2', 'Standard_F64s_v2', 'Standard_F72s_v2', 'Standard_FX24mds', 'Standard_FX36mds', 'Standard_FX48mds', 'Standard_E2s_v3', 'Standard_E4s_v3', 'Standard_E8s_v3', 'Standard_E16s_v3', 'Standard_E32s_v3', 'Standard_E48s_v3', 'Standard_E64s_v3', 'Standard_NC4as_T4_v3', 'Standard_NC6s_v3', 'Standard_NC8as_T4_v3', 'Standard_NC12s_v3', 'Standard_NC16as_T4_v3', 'Standard_NC24s_v3', 'Standard_NC64as_T4_v3', 'Standard_NC24ads_A100_v4', 'Standard_NC48ads_A100_v4', 'Standard_NC96ads_A100_v4', 'Standard_ND96asr_v4', 'Standard_ND96amsr_A100_v4', 'Standard_ND40rs_v2']
View in Studio: https://ml.azure.com/registries/azureml/models/distilbert-base-cased/version/13
License: apache-2.0
SharedComputeCapacityEnabled: True
SHA: 0dacbb01d604f8adeeb5b87c9339e485ac40d5c0
evaluation-min-sku-spec: 4|0|28|56
inference-min-sku-spec: 2|0|8|28
evaluation-recommended-sku: Standard_DS4_v2, Standard_D8a_v4, Standard_D8as_v4, Standard_DS5_v2, Standard_DS12_v2, Standard_D16a_v4, Standard_D16as_v4, Standard_D32a_v4, Standard_D32as_v4, Standard_D48a_v4, Standard_D48as_v4, Standard_D64a_v4, Standard_D64as_v4, Standard_D96a_v4, Standard_D96as_v4, Standard_FX4mds, Standard_FX12mds, Standard_F16s_v2, Standard_F32s_v2, Standard_F48s_v2, Standard_F64s_v2, Standard_F72s_v2, Standard_FX24mds, Standard_FX36mds, Standard_FX48mds, Standard_E4s_v3, Standard_E8s_v3, Standard_E16s_v3, Standard_E32s_v3, Standard_E48s_v3, Standard_E64s_v3, Standard_NC4as_T4_v3, Standard_NC6s_v3, Standard_NC8as_T4_v3, Standard_NC12s_v3, Standard_NC16as_T4_v3, Standard_NC24s_v3, Standard_NC64as_T4_v3, Standard_NC24ads_A100_v4, Standard_NC48ads_A100_v4, Standard_NC96ads_A100_v4, Standard_ND96asr_v4, Standard_ND96amsr_A100_v4, Standard_ND40rs_v2
inference-recommended-sku: Standard_DS3_v2, Standard_D4a_v4, Standard_D4as_v4, Standard_DS4_v2, Standard_D8a_v4, Standard_D8as_v4, Standard_DS5_v2, Standard_D16a_v4, Standard_D16as_v4, Standard_D32a_v4, Standard_D32as_v4, Standard_D48a_v4, Standard_D48as_v4, Standard_D64a_v4, Standard_D64as_v4, Standard_D96a_v4, Standard_D96as_v4, Standard_F4s_v2, Standard_FX4mds, Standard_F8s_v2, Standard_FX12mds, Standard_F16s_v2, Standard_F32s_v2, Standard_F48s_v2, Standard_F64s_v2, Standard_F72s_v2, Standard_FX24mds, Standard_FX36mds, Standard_FX48mds, Standard_E2s_v3, Standard_E4s_v3, Standard_E8s_v3, Standard_E16s_v3, Standard_E32s_v3, Standard_E48s_v3, Standard_E64s_v3, Standard_NC4as_T4_v3, Standard_NC6s_v3, Standard_NC8as_T4_v3, Standard_NC12s_v3, Standard_NC16as_T4_v3, Standard_NC24s_v3, Standard_NC64as_T4_v3, Standard_NC24ads_A100_v4, Standard_NC48ads_A100_v4, Standard_NC96ads_A100_v4, Standard_ND96asr_v4, Standard_ND96amsr_A100_v4, Standard_ND40rs_v2
languages: en