# Efficient training

A large language model fine-tuned from the gemma-2-9b-it model using the Weighted Preference Optimization (WPO) method, enhancing the effectiveness of off-policy preference optimization.

This is a 774M-parameter language model based on the GPT-2 architecture, trained on 150 billion tokens from the FineWeb dataset.

RWKV is a large language model that combines the advantages of RNN and Transformer, supporting efficient training and fast inference with unlimited context length processing capability.

Gerbil-A-32m is an A-grade model with 32 million parameters, trained on 640 million tokens, suitable for various natural language processing tasks.

DETA is a transformer-based object detection model that achieves rapid convergence and efficient detection by reintroducing the IoU assignment mechanism and NMS methods.

A German RoBERTa model trained using the WECHSEL method, achieving cross-lingual transfer of monolingual language models through effective initialization of subword embeddings.

A French version of GPT-2 trained using the WECHSEL method, achieving cross-lingual transfer of monolingual language models through effective initialization of subword embeddings.

A Chinese GPT-2 model trained using the WECHSEL method, achieving cross-lingual transfer of monolingual language models through effective initialization of subword embeddings.

This model is based on the BERT-tiny architecture and fine-tuned on the STS-B dataset using the M-FAC second-order optimizer for text similarity calculation.