license stringlengths 2 30 | tags stringlengths 2 513 | is_nc bool 1 class | readme_section stringlengths 201 597k | hash stringlengths 32 32 |
|---|---|---|---|---|
apache-2.0 | ['generated_from_trainer'] | false | tiny-mlm-glue-rte-target-glue-cola This model is a fine-tuned version of [muhtasham/tiny-mlm-glue-rte](https://huggingface.co/muhtasham/tiny-mlm-glue-rte) on the None dataset. It achieves the following results on the evaluation set: - Loss: 0.7986 - Matthews Correlation: 0.1168 | 4d5ca755b8bfb741503f9ae3cf9fc59f |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Matthews Correlation | |:-------------:|:-----:|:----:|:---------------:|:--------------------:| | 0.6097 | 1.87 | 500 | 0.6209 | 0.0 | | 0.6011 | 3.73 | 1000 | 0.6173 | 0.0 | | 0.5827 | 5.6 | 1500 | 0.6197 | 0.0622 | | 0.5534 | 7.46 | 2000 | 0.6410 | 0.0939 | | 0.5244 | 9.33 | 2500 | 0.6664 | 0.1184 | | 0.5087 | 11.19 | 3000 | 0.6684 | 0.1327 | | 0.4867 | 13.06 | 3500 | 0.6789 | 0.0999 | | 0.4693 | 14.93 | 4000 | 0.7124 | 0.1109 | | 0.4483 | 16.79 | 4500 | 0.7333 | 0.1388 | | 0.4303 | 18.66 | 5000 | 0.7486 | 0.1287 | | 0.4105 | 20.52 | 5500 | 0.7961 | 0.1321 | | 0.4046 | 22.39 | 6000 | 0.7986 | 0.1168 | | 4b05a25787011b5c1607d9c61855758f |
apache-2.0 | ['generated_from_trainer'] | false | wav2vec2-base-timit-demo-google-colab This model is a fine-tuned version of [facebook/wav2vec2-base](https://huggingface.co/facebook/wav2vec2-base) on the None dataset. It achieves the following results on the evaluation set: - Loss: 0.5090 - Wer: 0.3435 | c17a37badb3df266dc8e1582ca23e71b |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Wer | |:-------------:|:-----:|:-----:|:---------------:|:------:| | 3.5501 | 1.0 | 500 | 1.9752 | 0.9950 | | 0.8608 | 2.01 | 1000 | 0.5051 | 0.5035 | | 0.43 | 3.01 | 1500 | 0.4485 | 0.4525 | | 0.2921 | 4.02 | 2000 | 0.4658 | 0.4332 | | 0.2248 | 5.02 | 2500 | 0.4262 | 0.4268 | | 0.1863 | 6.02 | 3000 | 0.4126 | 0.3977 | | 0.1542 | 7.03 | 3500 | 0.4795 | 0.3987 | | 0.1374 | 8.03 | 4000 | 0.4882 | 0.3982 | | 0.1231 | 9.04 | 4500 | 0.4312 | 0.3790 | | 0.1082 | 10.04 | 5000 | 0.4344 | 0.3679 | | 0.0949 | 11.04 | 5500 | 0.4720 | 0.3769 | | 0.0897 | 12.05 | 6000 | 0.5382 | 0.3706 | | 0.0816 | 13.05 | 6500 | 0.4946 | 0.3618 | | 0.0726 | 14.06 | 7000 | 0.5383 | 0.3630 | | 0.0656 | 15.06 | 7500 | 0.4944 | 0.3693 | | 0.059 | 16.06 | 8000 | 0.5096 | 0.3639 | | 0.0572 | 17.07 | 8500 | 0.5066 | 0.3572 | | 0.0559 | 18.07 | 9000 | 0.5366 | 0.3610 | | 0.0468 | 19.08 | 9500 | 0.5103 | 0.3604 | | 0.0413 | 20.08 | 10000 | 0.5126 | 0.3496 | | 0.044 | 21.08 | 10500 | 0.5055 | 0.3524 | | 0.0351 | 22.09 | 11000 | 0.5526 | 0.3515 | | 0.0328 | 23.09 | 11500 | 0.4884 | 0.3512 | | 0.032 | 24.1 | 12000 | 0.5167 | 0.3474 | | 0.0271 | 25.1 | 12500 | 0.5027 | 0.3495 | | 0.0229 | 26.1 | 13000 | 0.5076 | 0.3444 | | 0.0252 | 27.11 | 13500 | 0.5122 | 0.3464 | | 0.0224 | 28.11 | 14000 | 0.5133 | 0.3447 | | 0.0236 | 29.12 | 14500 | 0.5090 | 0.3435 | | 28e9583b344c9f210122fd7236333936 |
apache-2.0 | ['automatic-speech-recognition', 'de'] | false | exp_w2v2r_de_vp-100k_age_teens-0_sixties-10_s50 Fine-tuned [facebook/wav2vec2-large-100k-voxpopuli](https://huggingface.co/facebook/wav2vec2-large-100k-voxpopuli) for speech recognition using the train split of [Common Voice 7.0 (de)](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0). When using this model, make sure that your speech input is sampled at 16kHz. This model has been fine-tuned by the [HuggingSound](https://github.com/jonatasgrosman/huggingsound) tool. | edc29df6e958eb80f5b37c4bc967a42b |
mit | [] | false | Artist_Yukiko Kanagai on Stable Diffusion This is the `<Yukiko Kanagai >` concept taught to Stable Diffusion via Textual Inversion. You can load this concept into the [Stable Conceptualizer](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_conceptualizer_inference.ipynb) notebook. You can also train your own concepts and load them into the concept libraries using [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb). Here is the new concept you will be able to use as a `style`:      | 64e0da67c64364f694103f90f5dd6026 |
mit | ['generated_from_trainer'] | false | xlm-roberta-base-finetuned-panx-en This model is a fine-tuned version of [xlm-roberta-base](https://huggingface.co/xlm-roberta-base) on the xtreme dataset. It achieves the following results on the evaluation set: - Loss: 0.4025 - F1: 0.6778 | 4e538ba4c46fe58260b8ba31a4c710c2 |
mit | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | F1 | |:-------------:|:-----:|:----:|:---------------:|:------:| | 1.1069 | 1.0 | 50 | 0.5201 | 0.5010 | | 0.4975 | 2.0 | 100 | 0.4503 | 0.6198 | | 0.3705 | 3.0 | 150 | 0.4025 | 0.6778 | | 14cd51481f582b31e3abf5e74ef65665 |
mit | [] | false | Description This model is a RoBERTa-based model pre-trained from scratch on Dutch hospital notes sourced from Electronic Health Records. The model is not fine-tuned. All code used for the creation of MedRoBERTa.nl can be found at https://github.com/cltl-students/verkijk_stella_rma_thesis_dutch_medical_language_model. | 88fb3bed2e693f4766d79341bf39dd40 |
mit | [] | false | Privacy By anonymizing the training data we made sure the model did not learn any representative associations linked to names. Apart from the training data, the model's vocabulary was also anonymized. This ensures that the model can not predict any names in the generative fill-mask task. | 0496fefd2176337fee46aedd575bd1a9 |
apache-2.0 | ['generated_from_trainer'] | false | finetuning-sentiment-model-3000-samples This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the imdb dataset. It achieves the following results on the evaluation set: - Loss: 0.3150 - Accuracy: 0.8633 - F1: 0.8656 | 5f8dc79f725fed85d3b2bcf7eaf674be |
cc-by-4.0 | ['questions and answers generation'] | false | Model Card of `lmqg/mt5-small-koquad-qag` This model is fine-tuned version of [google/mt5-small](https://huggingface.co/google/mt5-small) for question & answer pair generation task on the [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) (dataset_name: default) via [`lmqg`](https://github.com/asahi417/lm-question-generation). | 54068691efc3dda87095b4d4196e5ee8 |
cc-by-4.0 | ['questions and answers generation'] | false | Overview - **Language model:** [google/mt5-small](https://huggingface.co/google/mt5-small) - **Language:** ko - **Training data:** [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) (default) - **Online Demo:** [https://autoqg.net/](https://autoqg.net/) - **Repository:** [https://github.com/asahi417/lm-question-generation](https://github.com/asahi417/lm-question-generation) - **Paper:** [https://arxiv.org/abs/2210.03992](https://arxiv.org/abs/2210.03992) | 8c4d66666cde768f88109b4b913d733a |
cc-by-4.0 | ['questions and answers generation'] | false | model prediction question_answer_pairs = model.generate_qa("1990년 영화 《 남부군 》에서 단역으로 영화배우 첫 데뷔에 이어 같은 해 KBS 드라마 《지구인》에서 단역으로 출연하였고 이듬해 MBC 《여명의 눈동자》를 통해 단역으로 출연하였다.") ``` - With `transformers` ```python from transformers import pipeline pipe = pipeline("text2text-generation", "lmqg/mt5-small-koquad-qag") output = pipe("1990년 영화 《 남부군 》에서 단역으로 영화배우 첫 데뷔에 이어 같은 해 KBS 드라마 《지구인》에서 단역으로 출연하였고 이듬해 MBC 《여명의 눈동자》를 통해 단역으로 출연하였다.") ``` | 683d1439b3ab9d9bff66e40e17727215 |
cc-by-4.0 | ['questions and answers generation'] | false | Evaluation - ***Metric (Question & Answer Generation)***: [raw metric file](https://huggingface.co/lmqg/mt5-small-koquad-qag/raw/main/eval/metric.first.answer.paragraph.questions_answers.lmqg_qag_koquad.default.json) | | Score | Type | Dataset | |:--------------------------------|--------:|:--------|:-------------------------------------------------------------------| | QAAlignedF1Score (BERTScore) | 74.23 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | QAAlignedF1Score (MoverScore) | 75.06 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | QAAlignedPrecision (BERTScore) | 74.29 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | QAAlignedPrecision (MoverScore) | 75.14 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | QAAlignedRecall (BERTScore) | 74.2 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | QAAlignedRecall (MoverScore) | 75.04 | default | [lmqg/qag_koquad](https://huggingface.co/datasets/lmqg/qag_koquad) | | 7c8dd4d889bce17d1f2ea61dae182d26 |
cc-by-4.0 | ['questions and answers generation'] | false | Training hyperparameters The following hyperparameters were used during fine-tuning: - dataset_path: lmqg/qag_koquad - dataset_name: default - input_types: ['paragraph'] - output_types: ['questions_answers'] - prefix_types: None - model: google/mt5-small - max_length: 512 - max_length_output: 256 - epoch: 13 - batch: 8 - lr: 0.0005 - fp16: False - random_seed: 1 - gradient_accumulation_steps: 16 - label_smoothing: 0.0 The full configuration can be found at [fine-tuning config file](https://huggingface.co/lmqg/mt5-small-koquad-qag/raw/main/trainer_config.json). | c8e1f814c467f962b51e519595261e96 |
apache-2.0 | ['generated_from_trainer'] | false | vit-base-patch16-224-in21k-finetuned-emotion-classification-balanced-data-fer2013-affecthq-v0.0 This model is a fine-tuned version of [google/vit-base-patch16-224-in21k](https://huggingface.co/google/vit-base-patch16-224-in21k) on the None dataset. It achieves the following results on the evaluation set: - Loss: 1.0842 - Accuracy: 0.5958 | aa8e9ea5a9d5e43de93d381d05b56201 |
apache-2.0 | ['generated_from_trainer'] | false | Training hyperparameters The following hyperparameters were used during training: - learning_rate: 1e-05 - train_batch_size: 32 - eval_batch_size: 32 - seed: 17 - gradient_accumulation_steps: 4 - total_train_batch_size: 128 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: linear - lr_scheduler_warmup_ratio: 0.1 - num_epochs: 10 | f4c259cd99114d1ed6a25917f10cb27c |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Accuracy | |:-------------:|:-----:|:----:|:---------------:|:--------:| | 1.7482 | 1.0 | 133 | 1.7009 | 0.4323 | | 1.4052 | 2.0 | 266 | 1.3518 | 0.4998 | | 1.2372 | 3.0 | 399 | 1.2425 | 0.5344 | | 1.1663 | 4.0 | 532 | 1.1871 | 0.5494 | | 1.1238 | 5.0 | 665 | 1.1443 | 0.5817 | | 1.0124 | 6.0 | 798 | 1.1228 | 0.5869 | | 1.0262 | 7.0 | 931 | 1.1035 | 0.5920 | | 0.9963 | 8.0 | 1064 | 1.0917 | 0.5934 | | 0.9739 | 9.0 | 1197 | 1.0870 | 0.5948 | | 0.986 | 10.0 | 1330 | 1.0842 | 0.5958 | | a6adcf30958faf091afc413dfad3141f |
apache-2.0 | ['generated_from_trainer'] | false | distilbert-base-uncased-finetuned-cola This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the glue dataset. It achieves the following results on the evaluation set: - Loss: 0.4969 - Matthews Correlation: 0.4354 | b608cb5b99fad05ba883c21b703dd7ba |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Matthews Correlation | |:-------------:|:-----:|:----:|:---------------:|:--------------------:| | 0.5287 | 1.0 | 535 | 0.4969 | 0.4354 | | ee3a366b7dcad51935ae7e4fe183ed08 |
apache-2.0 | ['classical chinese', 'literary chinese', 'ancient chinese', 'masked-lm'] | false | Model Description This is a RoBERTa model pre-trained on Classical Chinese texts, derived from [GuwenBERT-base](https://huggingface.co/ethanyt/guwenbert-base). Character-embeddings are enhanced into traditional/simplified characters. You can fine-tune `roberta-classical-chinese-base-char` for downstream tasks, such as [sentence-segmentation](https://huggingface.co/KoichiYasuoka/roberta-classical-chinese-base-sentence-segmentation), [POS-tagging](https://huggingface.co/KoichiYasuoka/roberta-classical-chinese-base-upos), [dependency-parsing](https://huggingface.co/KoichiYasuoka/roberta-classical-chinese-base-ud-goeswith), and so on. | 8ce0b22a6abc94d55f88eadd3923b40b |
apache-2.0 | ['classical chinese', 'literary chinese', 'ancient chinese', 'masked-lm'] | false | How to Use ```py from transformers import AutoTokenizer,AutoModelForMaskedLM tokenizer=AutoTokenizer.from_pretrained("KoichiYasuoka/roberta-classical-chinese-base-char") model=AutoModelForMaskedLM.from_pretrained("KoichiYasuoka/roberta-classical-chinese-base-char") ``` | d30980c6a7cad4afbe694722b89fe771 |
apache-2.0 | [] | false | Model Description This **DAMO-YOLO-S** model is a small-size object detection model with fast inference speed and high accuracy, trained by **DAMO-YOLO**. DAMO-YOLO is a fast and accurate object detection method, which is developed by TinyML Team from Alibaba DAMO Data Analytics and Intelligence Lab. And it achieves a higher performance than state-of-the-art YOLO series. DAMO-YOLO is extend from YOLO but with some new techs, including Neural Architecture Search (NAS) backbones, efficient Reparameterized Generalized-FPN (RepGFPN), a lightweight head with AlignedOTA label assignment, and distillation enhancement. For more details, please refer to our [Arxiv Report](https://arxiv.org/abs/2211.15444) and [Github Code](https://github.com/tinyvision/DAMO-YOLO). Moreover, here you can find not only powerful models, but also highly efficient training strategies and complete tools from training to deployment. | f5a29e1cf3fe15442ae615b3d2986133 |
cc-by-4.0 | ['question generation'] | false | Model Card of `research-backup/bart-large-squadshifts-vanilla-reddit-qg` This model is fine-tuned version of [facebook/bart-large](https://huggingface.co/facebook/bart-large) for question generation task on the [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) (dataset_name: reddit) via [`lmqg`](https://github.com/asahi417/lm-question-generation). | e23f1498c1b7c7bf83b40bc5ae19d95f |
cc-by-4.0 | ['question generation'] | false | Overview - **Language model:** [facebook/bart-large](https://huggingface.co/facebook/bart-large) - **Language:** en - **Training data:** [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) (reddit) - **Online Demo:** [https://autoqg.net/](https://autoqg.net/) - **Repository:** [https://github.com/asahi417/lm-question-generation](https://github.com/asahi417/lm-question-generation) - **Paper:** [https://arxiv.org/abs/2210.03992](https://arxiv.org/abs/2210.03992) | 3d8d77f2a1d291b03990a45dc9c885b8 |
cc-by-4.0 | ['question generation'] | false | model prediction questions = model.generate_q(list_context="William Turner was an English painter who specialised in watercolour landscapes", list_answer="William Turner") ``` - With `transformers` ```python from transformers import pipeline pipe = pipeline("text2text-generation", "research-backup/bart-large-squadshifts-vanilla-reddit-qg") output = pipe("<hl> Beyonce <hl> further expanded her acting career, starring as blues singer Etta James in the 2008 musical biopic, Cadillac Records.") ``` | 21aff96ef78ce5d051e28f0fa48fe920 |
cc-by-4.0 | ['question generation'] | false | Evaluation - ***Metric (Question Generation)***: [raw metric file](https://huggingface.co/research-backup/bart-large-squadshifts-vanilla-reddit-qg/raw/main/eval/metric.first.sentence.paragraph_answer.question.lmqg_qg_squadshifts.reddit.json) | | Score | Type | Dataset | |:-----------|--------:|:-------|:---------------------------------------------------------------------------| | BERTScore | 92.19 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | Bleu_1 | 26.22 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | Bleu_2 | 16.98 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | Bleu_3 | 11.22 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | Bleu_4 | 7.74 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | METEOR | 20.72 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | MoverScore | 61.37 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | ROUGE_L | 24.81 | reddit | [lmqg/qg_squadshifts](https://huggingface.co/datasets/lmqg/qg_squadshifts) | | f1ac52610585731f3cd69f6439ac8d2d |
cc-by-4.0 | ['question generation'] | false | Training hyperparameters The following hyperparameters were used during fine-tuning: - dataset_path: lmqg/qg_squadshifts - dataset_name: reddit - input_types: ['paragraph_answer'] - output_types: ['question'] - prefix_types: None - model: facebook/bart-large - max_length: 512 - max_length_output: 32 - epoch: 2 - batch: 32 - lr: 0.0001 - fp16: False - random_seed: 1 - gradient_accumulation_steps: 2 - label_smoothing: 0.15 The full configuration can be found at [fine-tuning config file](https://huggingface.co/research-backup/bart-large-squadshifts-vanilla-reddit-qg/raw/main/trainer_config.json). | aa5d769995be7d5f5aec009a4c3a21d3 |
apache-2.0 | ['lexical normalization'] | false | Fine-tuned ByT5-small for MultiLexNorm (Serbian version)  This is the official release of the fine-tuned models for **the winning entry** to the [*W-NUT 2021: Multilingual Lexical Normalization (MultiLexNorm)* shared task](https://noisy-text.github.io/2021/multi-lexnorm.html), which evaluates lexical-normalization systems on 12 social media datasets in 11 languages. Our system is based on [ByT5](https://arxiv.org/abs/2105.13626), which we first pre-train on synthetic data and then fine-tune on authentic normalization data. It achieves the best performance by a wide margin in intrinsic evaluation, and also the best performance in extrinsic evaluation through dependency parsing. In addition to these fine-tuned models, we also release the source files on [GitHub](https://github.com/ufal/multilexnorm2021) and an interactive demo on [Google Colab](https://colab.research.google.com/drive/1rxpI8IlKk-D2crFqi2hdzbTBIezqgsCg?usp=sharing). | 9a5b4a7807e3c47434fd5a522e1526f0 |
mit | ['conversational'] | false | Chat with the model: ```python from transformers import AutoTokenizer, AutoModelWithLMHead tokenizer = AutoTokenizer.from_pretrained("r3dhummingbird/DialoGPT-medium-joshua") model = AutoModelWithLMHead.from_pretrained("r3dhummingbird/DialoGPT-medium-joshua") | c8254bd4854f2c087866becc5ea2cc69 |
mit | ['bart', 'pytorch'] | false | BART-IT - FanPage BART-IT is a sequence-to-sequence model, based on the BART architecture that is specifically tailored to the Italian language. The model is pre-trained on a [large corpus of Italian text](https://huggingface.co/datasets/gsarti/clean_mc4_it), and can be fine-tuned on a variety of tasks. | 40c5ff0387015c64be0945aebee69dc7 |
mit | ['bart', 'pytorch'] | false | Model description The model is a `base-`sized BART model, with a vocabulary size of 52,000 tokens. It has 140M parameters and can be used for any task that requires a sequence-to-sequence model. It is trained from scratch on a large corpus of Italian text, and can be fine-tuned on a variety of tasks. | 74a3f41227e349d57c733ba664bc88bc |
mit | ['bart', 'pytorch'] | false | Fine-tuning The model has been fine-tuned for the abstractive summarization task on 3 different Italian datasets: - **This model** [FanPage](https://huggingface.co/datasets/ARTeLab/fanpage) - finetuned model [here](https://huggingface.co/morenolq/bart-it-fanpage) - [IlPost](https://huggingface.co/datasets/ARTeLab/ilpost) - finetuned model [here](https://huggingface.co/morenolq/bart-it-ilpost) - [WITS](https://huggingface.co/datasets/Silvia/WITS) - finetuned model [here](https://huggingface.co/morenolq/bart-it-WITS) | 0432f3dea41fac4ec26cf4432158c518 |
mit | ['bart', 'pytorch'] | false | Usage In order to use the model, you can use the following code: ```python from transformers import AutoTokenizer, AutoModelForSeq2SeqLM tokenizer = AutoTokenizer.from_pretrained("morenolq/bart-it-fanpage") model = AutoModelForSeq2SeqLM.from_pretrained("morenolq/bart-it-fanpage") input_ids = tokenizer.encode("Il modello BART-IT è stato pre-addestrato su un corpus di testo italiano", return_tensors="pt") outputs = model.generate(input_ids, max_length=40, num_beams=4, early_stopping=True) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` | b8fd9fcdc0c2d8779690034a6978c322 |
mit | ['bart', 'pytorch'] | false | Citation If you find this model useful for your research, please cite the following paper: ```bibtex @Article{BARTIT, AUTHOR = {La Quatra, Moreno and Cagliero, Luca}, TITLE = {BART-IT: An Efficient Sequence-to-Sequence Model for Italian Text Summarization}, JOURNAL = {Future Internet}, VOLUME = {15}, YEAR = {2023}, NUMBER = {1}, ARTICLE-NUMBER = {15}, URL = {https://www.mdpi.com/1999-5903/15/1/15}, ISSN = {1999-5903}, DOI = {10.3390/fi15010015} } ``` | e17c9259f9ff3884cf5b2748a291e419 |
mit | [] | false | Generates Ad copy, currently for ads for Amazon shopping (fine tuned for electronics and wearables). **Usage Examples:** Enter the bolded text below to get the Amazon ad generated by the model. **Big savings on the new** Roku Streaming Device **Mothers Day discounts for** Apple Watch Wireless Charger USB Charging Cable **Big savings on the new Sony** **Last minute shopping for Samsung headphones for** You can try entering brand and product names like Samsung Galaxy to see the ad text generator in action. Currently fine tuned on the EleutherAI/gpt-neo-125M model **Model Performance:** The model does quite well on the Electronics and Wearables categories on which it has been fine-tuned. There are, however, occasional hallucinations, though the ad copy is mostly coherent. In other domains, it doesn't do quite as well... Tesla for Christmas today, Honda on sale | 6e7e14deee3e6aa51706dbbc05af4ce9 |
apache-2.0 | ['generated_from_trainer'] | false | Tagged_One_100v9_NER_Model_3Epochs_AUGMENTED This model is a fine-tuned version of [bert-base-cased](https://huggingface.co/bert-base-cased) on the tagged_one100v9_wikigold_split dataset. It achieves the following results on the evaluation set: - Loss: 0.4255 - Precision: 0.3040 - Recall: 0.2132 - F1: 0.2506 - Accuracy: 0.8539 | 0eeccd9f0408bc57af4ce1b8e09152ed |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Precision | Recall | F1 | Accuracy | |:-------------:|:-----:|:----:|:---------------:|:---------:|:------:|:------:|:--------:| | No log | 1.0 | 40 | 0.5167 | 0.1936 | 0.0376 | 0.0630 | 0.8004 | | No log | 2.0 | 80 | 0.4406 | 0.2405 | 0.1441 | 0.1802 | 0.8385 | | No log | 3.0 | 120 | 0.4255 | 0.3040 | 0.2132 | 0.2506 | 0.8539 | | bdf7e6a3952062e70e36fc782a808854 |
apache-2.0 | ['automatic-speech-recognition', 'pl'] | false | exp_w2v2t_pl_vp-sv_s571 Fine-tuned [facebook/wav2vec2-large-sv-voxpopuli](https://huggingface.co/facebook/wav2vec2-large-sv-voxpopuli) for speech recognition using the train split of [Common Voice 7.0 (pl)](https://huggingface.co/datasets/mozilla-foundation/common_voice_7_0). When using this model, make sure that your speech input is sampled at 16kHz. This model has been fine-tuned by the [HuggingSound](https://github.com/jonatasgrosman/huggingsound) tool. | 302a350bc4c1b6d592731e8751e5a717 |
apache-2.0 | [] | false | MobileBERT: a Compact Task-Agnostic BERT for Resource-Limited Devices MobileBERT is a thin version of BERT_LARGE, while equipped with bottleneck structures and a carefully designed balance between self-attentions and feed-forward networks. This checkpoint is the original MobileBert Optimized Uncased English: [uncased_L-24_H-128_B-512_A-4_F-4_OPT](https://storage.googleapis.com/cloud-tpu-checkpoints/mobilebert/uncased_L-24_H-128_B-512_A-4_F-4_OPT.tar.gz) checkpoint. | 478a5c3991aed98ebd9a5d779064ffb1 |
apache-2.0 | [] | false | How to use MobileBERT in `transformers` ```python from transformers import pipeline fill_mask = pipeline( "fill-mask", model="google/mobilebert-uncased", tokenizer="google/mobilebert-uncased" ) print( fill_mask(f"HuggingFace is creating a {fill_mask.tokenizer.mask_token} that the community uses to solve NLP tasks.") ) ``` | b6059ca5a8d63b204b0dcda12bdc1973 |
mit | [] | false | milady on Stable Diffusion This is the `<milady>` concept taught to Stable Diffusion via Textual Inversion. You can load this concept into the [Stable Conceptualizer](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_conceptualizer_inference.ipynb) notebook. You can also train your own concepts and load them into the concept libraries using [this notebook](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/sd_textual_inversion_training.ipynb). Here is the new concept you will be able to use as an `object`:     | 235a759dafe9ff5b13a7bae5e1fea8cd |
cc-by-sa-4.0 | [] | false | Model description This is a Japanese RoBERTa base model pre-trained on Japanese Wikipedia and the Japanese portion of CC-100. This model is trained with character-level tokenization and whole word masking. | 2d52303135ee86bac7f6939919783ef6 |
cc-by-sa-4.0 | [] | false | How to use You can use this model for masked language modeling as follows: ```python from transformers import AutoTokenizer, AutoModelForMaskedLM tokenizer = AutoTokenizer.from_pretrained('ku-nlp/roberta-base-japanese-char-wwm') model = AutoModelForMaskedLM.from_pretrained('ku-nlp/roberta-base-japanese-char-wwm') sentence = '京都大学で自然言語処理を[MASK]する。' encoding = tokenizer(sentence, return_tensors='pt') ... ``` You can fine-tune this model on downstream tasks. | d195f8a3c3686f4e42fecd53ba6652a8 |
cc-by-sa-4.0 | [] | false | Tokenization There is no need to tokenize texts in advance, and you can give raw texts to the tokenizer. The texts are tokenized into character-level tokens by [sentencepiece](https://github.com/google/sentencepiece). | 210ecf93c427dddcc40ef141ddd3ba4c |
cc-by-sa-4.0 | [] | false | Training procedure This model was trained on Japanese Wikipedia (as of 20220220) and the Japanese portion of CC-100. It took two weeks using 8 NVIDIA A100 GPUs. The following hyperparameters were used during pre-training: - learning_rate: 1e-4 - per_device_train_batch_size: 62 - distributed_type: multi-GPU - num_devices: 8 - gradient_accumulation_steps: 8 - total_train_batch_size: 3968 - max_seq_length: 512 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: linear schedule with warmup - training_steps: 330000 - warmup_steps: 10000 | e6fff935852f04d499107c211e477847 |
apache-2.0 | ['generated_from_trainer'] | false | distilbert-base-uncased-finetuned-emotion This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the emotion dataset. It achieves the following results on the evaluation set: - Loss: 0.2157 - Accuracy: 0.9265 - F1: 0.9267 | 298c9c76cb334707e807dbd98e29423f |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | |:-------------:|:-----:|:----:|:---------------:|:--------:|:------:| | 0.8322 | 1.0 | 250 | 0.3176 | 0.905 | 0.9015 | | 0.2481 | 2.0 | 500 | 0.2157 | 0.9265 | 0.9267 | | 4b6316e5e9ac9e5ff3d75c1bd6d47f83 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Please Note! This model is NOT the 19.2M images Characters Model on TrinArt, but an improved version of the original Trin-sama Twitter bot model. This model is intended to retain the original SD's aesthetics as much as possible while nudging the model to anime/manga style. Other TrinArt models can be found at: https://huggingface.co/naclbit/trinart_derrida_characters_v2_stable_diffusion https://huggingface.co/naclbit/trinart_characters_19.2m_stable_diffusion_v1 | fa18afdec6aa69c475d133957663db2a |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Diffusers The model has been ported to `diffusers` by [ayan4m1](https://huggingface.co/ayan4m1) and can easily be run from one of the branches: - `revision="diffusers-60k"` for the checkpoint trained on 60,000 steps, - `revision="diffusers-95k"` for the checkpoint trained on 95,000 steps, - `revision="diffusers-115k"` for the checkpoint trained on 115,000 steps. For more information, please have a look at [the "Three flavors" section]( | 201ac76a703710597625c977acfba0e5 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Gradio We also support a [Gradio](https://github.com/gradio-app/gradio) web ui with diffusers to run inside a colab notebook: [](https://colab.research.google.com/drive/1RWvik_C7nViiR9bNsu3fvMR3STx6RvDx?usp=sharing) | 6e8e5bdf5847fb85b76c69661dfff61e |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | using the 60,000 steps checkpoint pipe = StableDiffusionPipeline.from_pretrained("naclbit/trinart_stable_diffusion_v2", revision="diffusers-60k") pipe.to("cuda") image = pipe("A magical dragon flying in front of the Himalaya in manga style").images[0] image ```  If you want to run the pipeline faster or on a different hardware, please have a look at the [optimization docs](https://huggingface.co/docs/diffusers/optimization/fp16). | cb8fab7e390310aac95cb3a80fb944f0 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | !pip install diffusers==0.3.0 from diffusers import StableDiffusionImg2ImgPipeline import requests from PIL import Image from io import BytesIO url = "https://scitechdaily.com/images/Dog-Park.jpg" response = requests.get(url) init_image = Image.open(BytesIO(response.content)).convert("RGB") init_image = init_image.resize((768, 512)) | e9e1ce68cfd11633d2bbbdc40efff889 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | using the 115,000 steps checkpoint pipe = StableDiffusionImg2ImgPipeline.from_pretrained("naclbit/trinart_stable_diffusion_v2", revision="diffusers-115k") pipe.to("cuda") images = pipe(prompt="Manga drawing of Brad Pitt", init_image=init_image, strength=0.75, guidance_scale=7.5).images image ``` If you want to run the pipeline faster or on a different hardware, please have a look at the [optimization docs](https://huggingface.co/docs/diffusers/optimization/fp16). | d0dac82d8cc81613c154cff4a9f16397 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Stable Diffusion TrinArt/Trin-sama AI finetune v2 trinart_stable_diffusion is a SD model finetuned by about 40,000 assorted high resolution manga/anime-style pictures for 8 epochs. This is the same model running on Twitter bot @trinsama (https://twitter.com/trinsama) Twitterボット「とりんさまAI」@trinsama (https://twitter.com/trinsama) で使用しているSDのファインチューン済モデルです。一定のルールで選別された約4万枚のアニメ・マンガスタイルの高解像度画像を用いて約8エポックの訓練を行いました。 | 03bd093105c1f6aed844cdbf61087987 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Version 2 V2 checkpoint uses dropouts, 10,000 more images and a new tagging strategy and trained longer to improve results while retaining the original aesthetics. バージョン2は画像を1万枚追加したほか、ドロップアウトの適用、タグ付けの改善とより長いトレーニング時間により、SDのスタイルを保ったまま出力内容の改善を目指しています。 | 9abe8225cf934dd8a1fa6bfab59fc391 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Three flavors Step 115000/95000 checkpoints were trained further, but you may use step 60000 checkpoint instead if style nudging is too much. ステップ115000/95000のチェックポイントでスタイルが変わりすぎると感じる場合は、ステップ60000のチェックポイントを使用してみてください。 | 28e9711c73580e63689a3f70d0706d70 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | img2img If you want to run **latent-diffusion**'s stock ddim img2img script with this model, **use_ema** must be set to False. **latent-diffusion** のscriptsフォルダに入っているddim img2imgをこのモデルで動かす場合、use_emaはFalseにする必要があります。 | 6c43df7bf8d38d5dd0a976477c3da903 |
creativeml-openrail-m | ['stable-diffusion', 'stable-diffusion-diffusers', 'text-to-image'] | false | Examples Each images were diffused using K. Crowson's k-lms (from k-diffusion repo) method for 50 steps.    | 0ffc153bdd16ec42747161b2562eb3a7 |
mit | ['text2text-generation'] | false | Intro Trained on IndicNLGSuit [IndicQuestionGeneration](https://huggingface.co/datasets/ai4bharat/IndicQuestionGeneration) data for Bengali the model is finetuned from [IndicBART](https://huggingface.co/ai4bharat/IndicBART) | d8498765e942faa1f983500e5748195c |
mit | ['text2text-generation'] | false | Finetuned Command python run_summarization.py --model_name_or_path bnQG_models/checkpoint-32000 --do_eval --train_file train_bn.json --validation_file valid_bn.json --output_dir bnQG_models --overwrite_output_dir --per_device_train_batch_size=2 --per_device_eval_batch_size=4 --predict_with_generate --text_column src --summary_column tgt --save_steps 4000 --evaluation_strategy steps --gradient_accumulation_steps 4 --eval_steps 1000 --learning_rate 0.001 --num_beams 4 --forced_bos_token "<2bn>" --num_train_epochs 10 --warmup_steps 10000 | 6832f44c508357c01a2e61619d47a723 |
mit | ['text2text-generation'] | false | Inference script = "সুভাষ ১৮৯৭ খ্রিষ্টাব্দের ২৩ জানুয়ারি ব্রিটিশ ভারতের অন্তর্গত বাংলা প্রদেশের উড়িষ্যা বিভাগের (অধুনা, ভারতের ওড়িশা রাজ্য) কটকে জন্মগ্রহণ করেন।" answer = "১৮৯৭ খ্রিষ্টাব্দের ২৩ জানুয়ারি" inp = answer +" [SEP] "+script + " </s> <2bn>" inp_tok = tokenizer(inp, add_special_tokens=False, return_tensors="pt", padding=True).input_ids model.eval() | a8cd5e249d068bab555dc8db90abd70a |
mit | ['text2text-generation'] | false | Set dropouts to zero model_output=model.generate(inp_tok, use_cache=True, num_beams=4, max_length=20, min_length=1, early_stopping=True, pad_token_id=pad_id, bos_token_id=bos_id, eos_token_id=eos_id, decoder_start_token_id=tokenizer._convert_token_to_id_with_added_voc("<2bn>") ) decoded_output=tokenizer.decode(model_output[0], skip_special_tokens=True, clean_up_tokenization_spaces=False) | bfff15b33218a436a4f5c3227ea93b31 |
mit | ['text2text-generation'] | false | Citations @inproceedings{dabre2021indicbart, title={IndicBART: A Pre-trained Model for Natural Language Generation of Indic Languages}, author={Raj Dabre and Himani Shrotriya and Anoop Kunchukuttan and Ratish Puduppully and Mitesh M. Khapra and Pratyush Kumar}, year={2022}, booktitle={Findings of the Association for Computational Linguistics}, } @misc{kumar2022indicnlg, title={IndicNLG Suite: Multilingual Datasets for Diverse NLG Tasks in Indic Languages}, author={Aman Kumar and Himani Shrotriya and Prachi Sahu and Raj Dabre and Ratish Puduppully and Anoop Kunchukuttan and Amogh Mishra and Mitesh M. Khapra and Pratyush Kumar}, year={2022}, eprint={2203.05437}, archivePrefix={arXiv}, primaryClass={cs.CL} } | f315d464abda21996ed1ce48843fd496 |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity'] | false | all-mpnet-base-v2 This is a [sentence-transformers](https://www.SBERT.net) model: It maps sentences & paragraphs to a 768 dimensional dense vector space and can be used for tasks like clustering or semantic search. | e851893c05ee54a2d4c76f566c7941b5 |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity'] | false | Usage (Sentence-Transformers) Using this model becomes easy when you have [sentence-transformers](https://www.SBERT.net) installed: ``` pip install -U sentence-transformers ``` Then you can use the model like this: ```python from sentence_transformers import SentenceTransformer sentences = ["This is an example sentence", "Each sentence is converted"] model = SentenceTransformer('sentence-transformers/all-mpnet-base-v2') embeddings = model.encode(sentences) print(embeddings) ``` | 77dcdb9e1d64c686507c2de0a2d9573b |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity'] | false | Evaluation Results For an automated evaluation of this model, see the *Sentence Embeddings Benchmark*: [https://seb.sbert.net](https://seb.sbert.net?model_name=sentence-transformers/all-mpnet-base-v2) ------ | 5889da920c0679b2916cf4376779d241 |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity'] | false | Intended uses Our model is intented to be used as a sentence and short paragraph encoder. Given an input text, it ouptuts a vector which captures the semantic information. The sentence vector may be used for information retrieval, clustering or sentence similarity tasks. By default, input text longer than 384 word pieces is truncated. | d955afea4385d7608e979170fca25a8b |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity'] | false | Pre-training We use the pretrained [`microsoft/mpnet-base`](https://huggingface.co/microsoft/mpnet-base) model. Please refer to the model card for more detailed information about the pre-training procedure. | 8776c40ea4b90b75d9cb438b6d2a72ed |
apache-2.0 | ['vision', 'image-classification'] | false | Swin Transformer (large-sized model) Swin Transformer model trained on ImageNet-1k at resolution 224x224. It was introduced in the paper [Swin Transformer: Hierarchical Vision Transformer using Shifted Windows](https://arxiv.org/abs/2103.14030) by Liu et al. and first released in [this repository](https://github.com/microsoft/Swin-Transformer). Disclaimer: The team releasing Swin Transformer did not write a model card for this model so this model card has been written by the Hugging Face team. | 8caa513d98e1a7d9f843d982be2bd282 |
apache-2.0 | ['vision', 'image-classification'] | false | Model description The Swin Transformer is a type of Vision Transformer. It builds hierarchical feature maps by merging image patches (shown in gray) in deeper layers and has linear computation complexity to input image size due to computation of self-attention only within each local window (shown in red). It can thus serve as a general-purpose backbone for both image classification and dense recognition tasks. In contrast, previous vision Transformers produce feature maps of a single low resolution and have quadratic computation complexity to input image size due to computation of self-attention globally.  [Source](https://paperswithcode.com/method/swin-transformer) | d720cac1b86ab5ecdb76384a0454619d |
apache-2.0 | ['vision', 'image-classification'] | false | Intended uses & limitations You can use the raw model for image classification. See the [model hub](https://huggingface.co/models?search=swin) to look for fine-tuned versions on a task that interests you. | 715dea939e871ef3fa0e0c115fb016e4 |
apache-2.0 | ['vision', 'image-classification'] | false | How to use Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes: ```python from transformers import AutoFeatureExtractor, SwinForImageClassification from PIL import Image import requests url = "http://images.cocodataset.org/val2017/000000039769.jpg" image = Image.open(requests.get(url, stream=True).raw) feature_extractor = AutoFeatureExtractor.from_pretrained("microsoft/swin-large-patch4-window7-224") model = SwinForImageClassification.from_pretrained("microsoft/swin-large-patch4-window7-224") inputs = feature_extractor(images=image, return_tensors="pt") outputs = model(**inputs) logits = outputs.logits | 1af69650d4f5feed600dba2a9dfa051b |
apache-2.0 | ['vision', 'image-classification'] | false | model predicts one of the 1000 ImageNet classes predicted_class_idx = logits.argmax(-1).item() print("Predicted class:", model.config.id2label[predicted_class_idx]) ``` For more code examples, we refer to the [documentation](https://huggingface.co/transformers/model_doc/swin.html | 5a2d04760cbc4a5a3b5e37d42d268c28 |
apache-2.0 | ['vision', 'image-classification'] | false | BibTeX entry and citation info ```bibtex @article{DBLP:journals/corr/abs-2103-14030, author = {Ze Liu and Yutong Lin and Yue Cao and Han Hu and Yixuan Wei and Zheng Zhang and Stephen Lin and Baining Guo}, title = {Swin Transformer: Hierarchical Vision Transformer using Shifted Windows}, journal = {CoRR}, volume = {abs/2103.14030}, year = {2021}, url = {https://arxiv.org/abs/2103.14030}, eprinttype = {arXiv}, eprint = {2103.14030}, timestamp = {Thu, 08 Apr 2021 07:53:26 +0200}, biburl = {https://dblp.org/rec/journals/corr/abs-2103-14030.bib}, bibsource = {dblp computer science bibliography, https://dblp.org} } ``` | 3e98d9d24aefa2cc0e16c712a6e73d92 |
apache-2.0 | ['image-classification', 'vision'] | false | BEiT (large-sized model, fine-tuned on ImageNet-1k) BEiT model pre-trained in a self-supervised fashion on ImageNet-21k (14 million images, 21,841 classes) at resolution 224x224, and fine-tuned on ImageNet 2012 (1 million images, 1,000 classes) at resolution 224x224. It was introduced in the paper [BEIT: BERT Pre-Training of Image Transformers](https://arxiv.org/abs/2106.08254) by Hangbo Bao, Li Dong and Furu Wei and first released in [this repository](https://github.com/microsoft/unilm/tree/master/beit). Disclaimer: The team releasing BEiT did not write a model card for this model so this model card has been written by the Hugging Face team. | b110c5cf8a1baed27596eab409bf3ba2 |
apache-2.0 | ['image-classification', 'vision'] | false | How to use Here is how to use this model to classify an image of the COCO 2017 dataset into one of the 1,000 ImageNet classes: ```python from transformers import BeitFeatureExtractor, BeitForImageClassification from PIL import Image import requests url = 'http://images.cocodataset.org/val2017/000000039769.jpg' image = Image.open(requests.get(url, stream=True).raw) feature_extractor = BeitFeatureExtractor.from_pretrained('microsoft/beit-large-patch16-224') model = BeitForImageClassification.from_pretrained('microsoft/beit-large-patch16-224') inputs = feature_extractor(images=image, return_tensors="pt") outputs = model(**inputs) logits = outputs.logits | c901cecd063b174784df661660027132 |
apache-2.0 | ['translation'] | false | opus-mt-tum-es * source languages: tum * target languages: es * OPUS readme: [tum-es](https://github.com/Helsinki-NLP/OPUS-MT-train/blob/master/models/tum-es/README.md) * dataset: opus * model: transformer-align * pre-processing: normalization + SentencePiece * download original weights: [opus-2020-01-16.zip](https://object.pouta.csc.fi/OPUS-MT-models/tum-es/opus-2020-01-16.zip) * test set translations: [opus-2020-01-16.test.txt](https://object.pouta.csc.fi/OPUS-MT-models/tum-es/opus-2020-01-16.test.txt) * test set scores: [opus-2020-01-16.eval.txt](https://object.pouta.csc.fi/OPUS-MT-models/tum-es/opus-2020-01-16.eval.txt) | 14c4d1ee71d222bd25bba159bb4a08fa |
apache-2.0 | ['generated_from_keras_callback'] | false | toanbui1991/distilbert-base-uncased-finetuned-squad This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on an unknown dataset. It achieves the following results on the evaluation set: - Train Loss: 1.5101 - Train End Logits Accuracy: 0.6065 - Train Start Logits Accuracy: 0.5692 - Validation Loss: 1.1679 - Validation End Logits Accuracy: 0.6823 - Validation Start Logits Accuracy: 0.6523 - Epoch: 0 | 22014b7c1a21b568af7e4c1b47667be2 |
apache-2.0 | ['generated_from_keras_callback'] | false | Training results | Train Loss | Train End Logits Accuracy | Train Start Logits Accuracy | Validation Loss | Validation End Logits Accuracy | Validation Start Logits Accuracy | Epoch | |:----------:|:-------------------------:|:---------------------------:|:---------------:|:------------------------------:|:--------------------------------:|:-----:| | 1.5101 | 0.6065 | 0.5692 | 1.1679 | 0.6823 | 0.6523 | 0 | | 490bd6d1d3b2d6ee63c5faca8e6e8dfa |
apache-2.0 | ['generated_from_trainer'] | false | results This model is a fine-tuned version of [t5-small](https://huggingface.co/t5-small) on the None dataset. It achieves the following results on the evaluation set: - Loss: 0.2057 - Rouge2 Precision: 0.3564 - Rouge2 Recall: 0.2124 - Rouge2 Fmeasure: 0.256 | 049b9f291351f2841d4bfb1ec9fd70b5 |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Rouge2 Precision | Rouge2 Recall | Rouge2 Fmeasure | |:-------------:|:-----:|:----:|:---------------:|:----------------:|:-------------:|:---------------:| | No log | 1.0 | 240 | 0.3146 | 0.2121 | 0.1134 | 0.1424 | | No log | 2.0 | 480 | 0.2444 | 0.2855 | 0.1519 | 0.19 | | 0.6451 | 3.0 | 720 | 0.2195 | 0.3225 | 0.1821 | 0.223 | | 0.6451 | 4.0 | 960 | 0.2078 | 0.355 | 0.2113 | 0.2548 | | 0.2978 | 5.0 | 1200 | 0.2057 | 0.3564 | 0.2124 | 0.256 | | 5012b9f5e9a3ee0d5bb5de533cc0b0d8 |
apache-2.0 | [] | false | Cross-Encoder for Quora Duplicate Questions Detection This model was trained using [SentenceTransformers](https://sbert.net) [Cross-Encoder](https://www.sbert.net/examples/applications/cross-encoder/README.html) class. | 8a58642397ddba662c3faa430154977d |
apache-2.0 | [] | false | Training Data This model was trained on the [Quora Duplicate Questions](https://www.quora.com/q/quoradata/First-Quora-Dataset-Release-Question-Pairs) dataset. The model will predict a score between 0 and 1 how likely the two given questions are duplicates. Note: The model is not suitable to estimate the similarity of questions, e.g. the two questions "How to learn Java" and "How to learn Python" will result in a rahter low score, as these are not duplicates. | 9d86f658acf9fdb44cbdc770404703cc |
apache-2.0 | [] | false | Usage and Performance Pre-trained models can be used like this: ``` from sentence_transformers import CrossEncoder model = CrossEncoder('model_name') scores = model.predict([('Question 1', 'Question 2'), ('Question 3', 'Question 4')]) ``` You can use this model also without sentence_transformers and by just using Transformers ``AutoModel`` class | be77f7e2a82bc91c414c6606aae07ab3 |
apache-2.0 | ['generated_from_trainer'] | false | small-mlm-glue-mnli-custom-tokenizer This model is a fine-tuned version of [google/bert_uncased_L-4_H-512_A-8](https://huggingface.co/google/bert_uncased_L-4_H-512_A-8) on the None dataset. It achieves the following results on the evaluation set: - Loss: 5.6551 | cbbf818a9a15b2d15817555d6975109d |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | |:-------------:|:-----:|:-----:|:---------------:| | 7.0308 | 0.4 | 500 | 6.6001 | | 6.346 | 0.8 | 1000 | 6.3998 | | 6.1061 | 1.2 | 1500 | 6.3170 | | 5.9586 | 1.6 | 2000 | 6.2799 | | 5.8773 | 2.0 | 2500 | 6.2034 | | 5.7403 | 2.4 | 3000 | 6.1609 | | 5.6602 | 2.8 | 3500 | 6.1113 | | 5.5809 | 3.2 | 4000 | 6.1267 | | 5.5663 | 3.6 | 4500 | 6.0647 | | 5.6266 | 4.0 | 5000 | 6.1090 | | 5.4756 | 4.4 | 5500 | 6.0302 | | 5.4905 | 4.8 | 6000 | 6.0292 | | 5.3179 | 5.2 | 6500 | 5.9758 | | 5.3375 | 5.6 | 7000 | 6.0125 | | 5.3035 | 6.0 | 7500 | 5.9495 | | 5.1918 | 6.4 | 8000 | 5.9537 | | 5.2499 | 6.8 | 8500 | 5.9100 | | 5.1905 | 7.2 | 9000 | 5.8620 | | 5.1787 | 7.6 | 9500 | 5.9296 | | 5.1534 | 8.0 | 10000 | 5.9442 | | 5.1396 | 8.4 | 10500 | 5.8609 | | 5.1272 | 8.8 | 11000 | 5.8358 | | 4.9615 | 9.2 | 11500 | 5.8617 | | 5.0062 | 9.6 | 12000 | 5.8043 | | 5.0131 | 10.0 | 12500 | 5.8119 | | 4.9326 | 10.4 | 13000 | 5.7851 | | 4.9655 | 10.8 | 13500 | 5.7792 | | 4.9256 | 11.2 | 14000 | 5.7843 | | 4.9195 | 11.6 | 14500 | 5.7652 | | 4.8299 | 12.0 | 15000 | 5.7606 | | 4.8748 | 12.4 | 15500 | 5.7577 | | 4.7588 | 12.8 | 16000 | 5.7048 | | 4.8185 | 13.2 | 16500 | 5.7245 | | 4.7679 | 13.6 | 17000 | 5.7402 | | 4.7377 | 14.0 | 17500 | 5.7034 | | 4.7403 | 14.4 | 18000 | 5.7054 | | 4.6628 | 14.8 | 18500 | 5.7203 | | 4.6801 | 15.2 | 19000 | 5.6798 | | 4.6014 | 15.6 | 19500 | 5.6931 | | 4.618 | 16.0 | 20000 | 5.6620 | | 4.6037 | 16.4 | 20500 | 5.6441 | | 4.6004 | 16.8 | 21000 | 5.6262 | | 4.5432 | 17.2 | 21500 | 5.6726 | | 4.576 | 17.6 | 22000 | 5.6322 | | 4.5568 | 18.0 | 22500 | 5.6551 | | 3b497a39c1d1e10ce15034ed58bc1cfd |
apache-2.0 | ['generated_from_trainer'] | false | distilbert-base-uncased-finetuned-emotion This model is a fine-tuned version of [distilbert-base-uncased](https://huggingface.co/distilbert-base-uncased) on the emotion dataset. It achieves the following results on the evaluation set: - Loss: 0.1591 - Accuracy: 0.939 - F1: 0.9391 | b041e7703678b6bd4545002222a25505 |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Accuracy | F1 | |:-------------:|:-----:|:----:|:---------------:|:--------:|:------:| | 0.2497 | 1.0 | 1000 | 0.2133 | 0.9255 | 0.9252 | | 0.1498 | 2.0 | 2000 | 0.1652 | 0.934 | 0.9339 | | 0.0965 | 3.0 | 3000 | 0.1591 | 0.939 | 0.9391 | | 17a70039718ff39c806bbfd49722cbb1 |
apache-2.0 | ['generated_from_trainer'] | false | opus-mt-ar-en-finetuned-ar-to-en This model is a fine-tuned version of [Helsinki-NLP/opus-mt-ar-en](https://huggingface.co/Helsinki-NLP/opus-mt-ar-en) on the opus_infopankki dataset. It achieves the following results on the evaluation set: - Loss: 0.7269 - Bleu: 51.6508 - Gen Len: 15.0812 | ca00082c15cbb79875b0104f54573d63 |
apache-2.0 | ['generated_from_trainer'] | false | Training hyperparameters The following hyperparameters were used during training: - learning_rate: 2e-06 - train_batch_size: 32 - eval_batch_size: 32 - seed: 42 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: linear - num_epochs: 20 - mixed_precision_training: Native AMP | f62226925ccabce1cec78704fada0035 |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Bleu | Gen Len | |:-------------:|:-----:|:-----:|:---------------:|:-------:|:-------:| | 1.4974 | 1.0 | 1587 | 1.3365 | 36.9061 | 15.3385 | | 1.3768 | 2.0 | 3174 | 1.2139 | 39.5476 | 15.2079 | | 1.2887 | 3.0 | 4761 | 1.1265 | 41.2771 | 15.2034 | | 1.2076 | 4.0 | 6348 | 1.0556 | 42.6907 | 15.2687 | | 1.1512 | 5.0 | 7935 | 0.9975 | 43.9498 | 15.2072 | | 1.0797 | 6.0 | 9522 | 0.9491 | 45.224 | 15.2034 | | 1.0499 | 7.0 | 11109 | 0.9101 | 46.1387 | 15.1651 | | 1.0095 | 8.0 | 12696 | 0.8778 | 47.0586 | 15.1788 | | 0.9833 | 9.0 | 14283 | 0.8501 | 47.8083 | 15.162 | | 0.9601 | 10.0 | 15870 | 0.8267 | 48.5236 | 15.1784 | | 0.9457 | 11.0 | 17457 | 0.8059 | 49.1717 | 15.095 | | 0.9233 | 12.0 | 19044 | 0.7883 | 49.7742 | 15.1126 | | 0.8964 | 13.0 | 20631 | 0.7736 | 50.2168 | 15.0917 | | 0.8849 | 14.0 | 22218 | 0.7606 | 50.5583 | 15.0913 | | 0.8751 | 15.0 | 23805 | 0.7504 | 50.8481 | 15.1108 | | 0.858 | 16.0 | 25392 | 0.7417 | 51.1841 | 15.0989 | | 0.8673 | 17.0 | 26979 | 0.7353 | 51.4271 | 15.0939 | | 0.8548 | 18.0 | 28566 | 0.7306 | 51.535 | 15.0911 | | 0.8483 | 19.0 | 30153 | 0.7279 | 51.6102 | 15.078 | | 0.8614 | 20.0 | 31740 | 0.7269 | 51.6508 | 15.0812 | | 9bcc8af3269181c958d0456c09e2de9c |
cc-by-4.0 | ['espnet', 'audio', 'text-to-speech'] | false | Demo: How to use in ESPnet2 ```bash cd espnet git checkout 49a284e69308d81c142b89795de255b4ce290c54 pip install -e . cd egs2/talromur/tts1 ./run.sh --skip_data_prep false --skip_train true --download_model espnet/GunnarThor_talromur_g_fastspeech2 ``` | 9ebe75fb6b720b0aae02d8c06e3aacd7 |
cc-by-4.0 | ['espnet', 'audio', 'text-to-speech'] | false | TTS config <details><summary>expand</summary> ``` config: conf/tuning/train_fastspeech2.yaml print_config: false log_level: INFO dry_run: false iterator_type: sequence output_dir: exp/g/tts_train_fastspeech2_raw_phn_none ngpu: 1 seed: 0 num_workers: 1 num_att_plot: 3 dist_backend: nccl dist_init_method: env:// dist_world_size: null dist_rank: null local_rank: 0 dist_master_addr: null dist_master_port: null dist_launcher: null multiprocessing_distributed: false unused_parameters: false sharded_ddp: false cudnn_enabled: true cudnn_benchmark: false cudnn_deterministic: true collect_stats: false write_collected_feats: false max_epoch: 100 patience: null val_scheduler_criterion: - valid - loss early_stopping_criterion: - valid - loss - min best_model_criterion: - - valid - loss - min - - train - loss - min keep_nbest_models: 5 nbest_averaging_interval: 0 grad_clip: 1.0 grad_clip_type: 2.0 grad_noise: false accum_grad: 8 no_forward_run: false resume: true train_dtype: float32 use_amp: false log_interval: null use_matplotlib: true use_tensorboard: true use_wandb: false wandb_project: null wandb_id: null wandb_entity: null wandb_name: null wandb_model_log_interval: -1 detect_anomaly: false pretrain_path: null init_param: [] ignore_init_mismatch: false freeze_param: [] num_iters_per_epoch: 800 batch_size: 20 valid_batch_size: null batch_bins: 2500000 valid_batch_bins: null train_shape_file: - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/train/text_shape.phn - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/train/speech_shape valid_shape_file: - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/valid/text_shape.phn - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/valid/speech_shape batch_type: numel valid_batch_type: null fold_length: - 150 - 204800 sort_in_batch: descending sort_batch: descending multiple_iterator: false chunk_length: 500 chunk_shift_ratio: 0.5 num_cache_chunks: 1024 train_data_path_and_name_and_type: - - dump/raw/train_g_phn/text - text - text - - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/train_g_phn/durations - durations - text_int - - dump/raw/train_g_phn/wav.scp - speech - sound valid_data_path_and_name_and_type: - - dump/raw/dev_g_phn/text - text - text - - exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/dev_g_phn/durations - durations - text_int - - dump/raw/dev_g_phn/wav.scp - speech - sound allow_variable_data_keys: false max_cache_size: 0.0 max_cache_fd: 32 valid_max_cache_size: null optim: adam optim_conf: lr: 1.0 scheduler: noamlr scheduler_conf: model_size: 384 warmup_steps: 4000 token_list: - <blank> - <unk> - ',' - . - r - t - n - a0 - s - I0 - D - l - Y0 - m - v - h - E1 - k - a:1 - E:1 - f - G - j - T - a1 - p - c - au:1 - i:1 - O:1 - I:1 - E0 - I1 - r_0 - t_h - k_h - Y1 - ei1 - i0 - ou:1 - ei:1 - u:1 - O1 - N - l_0 - '91' - ai0 - au1 - ou0 - n_0 - ei0 - O0 - ou1 - ai:1 - '9:1' - ai1 - i1 - '90' - au0 - c_h - x - 9i:1 - C - p_h - u0 - Y:1 - J - 9i1 - u1 - 9i0 - N_0 - m_0 - J_0 - Oi1 - Yi0 - Yi1 - Oi0 - au:0 - '9:0' - E:0 - <sos/eos> odim: null model_conf: {} use_preprocessor: true token_type: phn bpemodel: null non_linguistic_symbols: null cleaner: null g2p: null feats_extract: fbank feats_extract_conf: n_fft: 1024 hop_length: 256 win_length: null fs: 22050 fmin: 80 fmax: 7600 n_mels: 80 normalize: global_mvn normalize_conf: stats_file: exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/train/feats_stats.npz tts: fastspeech2 tts_conf: adim: 384 aheads: 2 elayers: 4 eunits: 1536 dlayers: 4 dunits: 1536 positionwise_layer_type: conv1d positionwise_conv_kernel_size: 3 duration_predictor_layers: 2 duration_predictor_chans: 256 duration_predictor_kernel_size: 3 postnet_layers: 5 postnet_filts: 5 postnet_chans: 256 use_masking: true use_scaled_pos_enc: true encoder_normalize_before: true decoder_normalize_before: true reduction_factor: 1 init_type: xavier_uniform init_enc_alpha: 1.0 init_dec_alpha: 1.0 transformer_enc_dropout_rate: 0.2 transformer_enc_positional_dropout_rate: 0.2 transformer_enc_attn_dropout_rate: 0.2 transformer_dec_dropout_rate: 0.2 transformer_dec_positional_dropout_rate: 0.2 transformer_dec_attn_dropout_rate: 0.2 pitch_predictor_layers: 5 pitch_predictor_chans: 256 pitch_predictor_kernel_size: 5 pitch_predictor_dropout: 0.5 pitch_embed_kernel_size: 1 pitch_embed_dropout: 0.0 stop_gradient_from_pitch_predictor: true energy_predictor_layers: 2 energy_predictor_chans: 256 energy_predictor_kernel_size: 3 energy_predictor_dropout: 0.5 energy_embed_kernel_size: 1 energy_embed_dropout: 0.0 stop_gradient_from_energy_predictor: false pitch_extract: dio pitch_extract_conf: fs: 22050 n_fft: 1024 hop_length: 256 f0max: 400 f0min: 80 reduction_factor: 1 pitch_normalize: global_mvn pitch_normalize_conf: stats_file: exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/train/pitch_stats.npz energy_extract: energy energy_extract_conf: fs: 22050 n_fft: 1024 hop_length: 256 win_length: null reduction_factor: 1 energy_normalize: global_mvn energy_normalize_conf: stats_file: exp/g/tts_train_tacotron2_raw_phn_none/decode_use_teacher_forcingtrue_train.loss.ave/stats/train/energy_stats.npz required: - output_dir - token_list version: 0.10.7a1 distributed: false ``` </details> | 9dd334c4910fb60b67366d9e8064507b |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity', 'transformers'] | false | Sentence Transformers ```python from sentence_transformers import SentenceTransformer question = "<Q>How many models can I host on HuggingFace?" answer_1 = "<A>All plans come with unlimited private models and datasets." answer_2 = "<A>AutoNLP is an automatic way to train and deploy state-of-the-art NLP models, seamlessly integrated with the Hugging Face ecosystem." answer_3 = "<A>Based on how much training data and model variants are created, we send you a compute cost and payment link - as low as $10 per job." model = SentenceTransformer('clips/mfaq') embeddings = model.encode([question, answer_1, answer_3, answer_3]) print(embeddings) ``` | 20d9426ac5fee14fc2700e65622f8abf |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity', 'transformers'] | false | First element of model_output contains all token embeddings input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float() return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9) question = "<Q>How many models can I host on HuggingFace?" answer_1 = "<A>All plans come with unlimited private models and datasets." answer_2 = "<A>AutoNLP is an automatic way to train and deploy state-of-the-art NLP models, seamlessly integrated with the Hugging Face ecosystem." answer_3 = "<A>Based on how much training data and model variants are created, we send you a compute cost and payment link - as low as $10 per job." tokenizer = AutoTokenizer.from_pretrained('clips/mfaq') model = AutoModel.from_pretrained('clips/mfaq') | 4472d3bd61c842796c785a9c5d0a6592 |
apache-2.0 | ['sentence-transformers', 'feature-extraction', 'sentence-similarity', 'transformers'] | false | Citation information ``` @misc{debruyn2021mfaq, title={MFAQ: a Multilingual FAQ Dataset}, author={Maxime De Bruyn and Ehsan Lotfi and Jeska Buhmann and Walter Daelemans}, year={2021}, eprint={2109.12870}, archivePrefix={arXiv}, primaryClass={cs.CL} } ``` | 6e86a70b1664909af636ae33ee542229 |
apache-2.0 | ['generated_from_trainer'] | false | albert-large-v2-finetuned-wnli This model is a fine-tuned version of [albert-large-v2](https://huggingface.co/albert-large-v2) on the glue dataset. It achieves the following results on the evaluation set: - Loss: 0.6919 - Accuracy: 0.5352 | 36919be99a8dad7c6dc62ae38d19fd90 |
apache-2.0 | ['generated_from_trainer'] | false | Training results | Training Loss | Epoch | Step | Validation Loss | Accuracy | |:-------------:|:-----:|:----:|:---------------:|:--------:| | No log | 1.0 | 17 | 0.7292 | 0.4366 | | No log | 2.0 | 34 | 0.6919 | 0.5352 | | No log | 3.0 | 51 | 0.7084 | 0.4648 | | No log | 4.0 | 68 | 0.7152 | 0.5352 | | No log | 5.0 | 85 | 0.7343 | 0.5211 | | be330d669d7249b4887fae60715b7359 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.