# Multi-source data training

A convolutional neural network (CNN) built on the Keras framework, specifically designed to recognize individual Japanese characters from 64×64 grayscale images, supporting both handwritten and printed text recognition.

Uzbek speech recognition model fine-tuned based on Whisper medium, supports Tashkent dialect, trained on approximately 700 hours of data

kazRush-kk-ru is a Kazakh-to-Russian translation model based on the T5 configuration, trained on multiple parallel datasets.

The Skywork Critic series of models are advanced judgment models that excel in paired preference evaluation. They can compare and evaluate a pair of input contents and provide detailed judgments.

A text classification model fine-tuned based on GPT-2, used to distinguish AI-generated text, Zhihu user responses, and texts from other sources.

Real3D is a 2D-to-3D mapping Transformer model based on the TripoSR architecture, extending its capability to process real-world images through unsupervised self-training and automatic data filtering.

A Turkish text generation model fully fine-tuned on 30GB of Turkish dataset based on LLaMA-3 8B

A distilled automatic speech recognition model based on the Whisper architecture, optimized for Thai language with balanced performance and efficiency

A Russian text normalization model fine-tuned from FRED-T5-large, supporting normalization of numerals and Latin characters

TituLM-1B-BN-V1 is a large language model specifically trained for generating and understanding Bengali text, extensively trained on a dataset containing 4.51 billion Bengali tokens.

Hamsa is an Arabic speech recognition model built on the Whisper architecture, focusing on the linguistic needs of the Middle East and North Africa region.

The optimal version in the UniNER series, integrating named entity recognition models from three major data sources

Transformer-based French printed text OCR model, filling the gap of French version in TrOCR models

A bidirectional encoder model pre-trained for Russian language, trained on large-scale text corpora using standard masked language modeling objectives

Ultra-high-precision multi-class image classifier for accurate sensitive content detection

A Dutch text classification model based on BERT architecture for detecting sarcasm in news headlines

Swedish BERT base model released by the National Library of Sweden/KBLab, trained on multi-source texts

A model pre-trained on Bulgarian language using Masked Language Modeling (MLM) objective, case-sensitive.

Finnish speech recognition model fine-tuned based on XLSR-53 large model, supports 16kHz audio input

A Catalan speech recognition model fine-tuned based on facebook/wav2vec2-large-100k-voxpopuli

AlephBERT is a cutting-edge language model for Hebrew, based on Google's BERT architecture, specifically designed for processing Hebrew text.

Catalan automatic speech recognition model fine-tuned based on facebook/wav2vec2-large-xlsr-53

The state-of-the-art Hebrew language model based on the BERT architecture

Turkish automatic speech recognition model fine-tuned based on facebook/wav2vec2-xls-r-300m

This is a streamlined version of the Bulgarian RoBERTa model, containing only 6 hidden layers while maintaining comparable performance.

Russian toxicity text classifier based on rubert-base-cased-conversational model

A Swedish Bert Large model implemented based on the Megatron-LM framework, containing 340 million parameters, pre-trained on 85GB of Swedish text

Swedish BERT base model released by the National Library of Sweden/KBLab, trained on multi-source text data

The first BERT model specifically designed for the Moroccan Arabic dialect 'Darija', based on the BERT-base architecture, trained on approximately 3 million Darija dialect text sequences.

Finnish language model pre-trained on GPT-2 architecture, 117M parameter version

A DistilBERT fine-tuned model for English text punctuation restoration, specifically designed to add punctuation to lowercase English text without punctuation.

A compact Hebrew text generation model based on GPT2 architecture, trained on TPU and GPU