README.md
| 1 | --- |
| 2 | language: en |
| 3 | pipeline_tag: zero-shot-classification |
| 4 | tags: |
| 5 | - distilbert |
| 6 | datasets: |
| 7 | - multi_nli |
| 8 | metrics: |
| 9 | - accuracy |
| 10 | --- |
| 11 | |
| 12 | # DistilBERT base model (uncased) |
| 13 | |
| 14 | |
| 15 | ## Table of Contents |
| 16 | - [Model Details](#model-details) |
| 17 | - [How to Get Started With the Model](#how-to-get-started-with-the-model) |
| 18 | - [Uses](#uses) |
| 19 | - [Risks, Limitations and Biases](#risks-limitations-and-biases) |
| 20 | - [Training](#training) |
| 21 | - [Evaluation](#evaluation) |
| 22 | - [Environmental Impact](#environmental-impact) |
| 23 | |
| 24 | |
| 25 | |
| 26 | ## Model Details |
| 27 | **Model Description:** This is the [uncased DistilBERT model](https://huggingface.co/distilbert-base-uncased) fine-tuned on [Multi-Genre Natural Language Inference](https://huggingface.co/datasets/multi_nli) (MNLI) dataset for the zero-shot classification task. |
| 28 | - **Developed by:** The [Typeform](https://www.typeform.com/) team. |
| 29 | - **Model Type:** Zero-Shot Classification |
| 30 | - **Language(s):** English |
| 31 | - **License:** Unknown |
| 32 | - **Parent Model:** See the [distilbert base uncased model](https://huggingface.co/distilbert-base-uncased) for more information about the Distilled-BERT base model. |
| 33 | |
| 34 | |
| 35 | ## How to Get Started with the Model |
| 36 | |
| 37 | ```python |
| 38 | from transformers import AutoTokenizer, AutoModelForSequenceClassification |
| 39 | |
| 40 | tokenizer = AutoTokenizer.from_pretrained("typeform/distilbert-base-uncased-mnli") |
| 41 | |
| 42 | model = AutoModelForSequenceClassification.from_pretrained("typeform/distilbert-base-uncased-mnli") |
| 43 | |
| 44 | ``` |
| 45 | |
| 46 | ## Uses |
| 47 | This model can be used for text classification tasks. |
| 48 | |
| 49 | |
| 50 | ## Risks, Limitations and Biases |
| 51 | **CONTENT WARNING: Readers should be aware this section contains content that is disturbing, offensive, and can propagate historical and current stereotypes.** |
| 52 | |
| 53 | Significant research has explored bias and fairness issues with language models (see, e.g., [Sheng et al. (2021)](https://aclanthology.org/2021.acl-long.330.pdf) and [Bender et al. (2021)](https://dl.acm.org/doi/pdf/10.1145/3442188.3445922)). |
| 54 | |
| 55 | |
| 56 | ## Training |
| 57 | |
| 58 | #### Training Data |
| 59 | |
| 60 | |
| 61 | This model of DistilBERT-uncased is pretrained on the Multi-Genre Natural Language Inference [(MultiNLI)](https://huggingface.co/datasets/multi_nli) corpus. It is a crowd-sourced collection of 433k sentence pairs annotated with textual entailment information. The corpus covers a range of genres of spoken and written text, and supports a distinctive cross-genre generalization evaluation. |
| 62 | |
| 63 | This model is also **not** case-sensitive, i.e., it does not make a difference between "english" and "English". |
| 64 | |
| 65 | |
| 66 | #### Training Procedure |
| 67 | |
| 68 | Training is done on a [p3.2xlarge](https://aws.amazon.com/ec2/instance-types/p3/) AWS EC2 with the following hyperparameters: |
| 69 | |
| 70 | ``` |
| 71 | $ run_glue.py \ |
| 72 | --model_name_or_path distilbert-base-uncased \ |
| 73 | --task_name mnli \ |
| 74 | --do_train \ |
| 75 | --do_eval \ |
| 76 | --max_seq_length 128 \ |
| 77 | --per_device_train_batch_size 16 \ |
| 78 | --learning_rate 2e-5 \ |
| 79 | --num_train_epochs 5 \ |
| 80 | --output_dir /tmp/distilbert-base-uncased_mnli/ |
| 81 | ``` |
| 82 | |
| 83 | ## Evaluation |
| 84 | |
| 85 | |
| 86 | #### Evaluation Results |
| 87 | When fine-tuned on downstream tasks, this model achieves the following results: |
| 88 | - **Epoch = ** 5.0 |
| 89 | - **Evaluation Accuracy =** 0.8206875508543532 |
| 90 | - **Evaluation Loss =** 0.8706700205802917 |
| 91 | - ** Evaluation Runtime = ** 17.8278 |
| 92 | - ** Evaluation Samples per second = ** 551.498 |
| 93 | |
| 94 | MNLI and MNLI-mm results: |
| 95 | |
| 96 | | Task | MNLI | MNLI-mm | |
| 97 | |:----:|:----:|:----:| |
| 98 | | | 82.0 | 82.0 | |
| 99 | |
| 100 | |
| 101 | |
| 102 | ## Environmental Impact |
| 103 | |
| 104 | Carbon emissions can be estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). We present the hardware type based on the [associated paper](https://arxiv.org/pdf/2105.09680.pdf). |
| 105 | |
| 106 | |
| 107 | **Hardware Type:** 1 NVIDIA Tesla V100 GPUs |
| 108 | |
| 109 | **Hours used:** Unknown |
| 110 | |
| 111 | **Cloud Provider:** AWS EC2 P3 |
| 112 | |
| 113 | |
| 114 | **Compute Region:** Unknown |
| 115 | |
| 116 | |
| 117 | |
| 118 | **Carbon Emitted:** (Power consumption x Time x Carbon produced based on location of power grid): Unknown |
| 119 | |
| 120 | |