# Efficient Training

nanoVLM is a lightweight vision-language model (VLM) designed for efficient training and experimentation.

nanoVLM is a lightweight vision-language model (VLM) designed for efficient training and experimentation.

A text generation and inference model optimized using Unsloth and TRL libraries based on the Qwen/Qwen2.5-Coder-7B model, achieving 2x faster training speed

Bonsai is a small ternary-weighted language model with 500 million parameters, built on the Llama architecture and using the Mistral tokenizer, trained on fewer than 5 billion tokens.

RWKV-7 model using Flash Linear Attention format, trained on the Pile dataset, supporting English text generation tasks.

TraceBack 12b is a 4bit quantized version based on the Mistral-Nemo-Instruct architecture, focusing on instruction-following and chain-of-thought reasoning tasks.

This is a speech language model based on discrete Hubert tokens, focusing on efficient training and capable of generating speech segment continuations.

Open Reasoner Zero is an open-source solution for large-scale reinforcement learning based on foundational models, focusing on scalability, simplicity, and ease of use for large-scale reasoning-oriented reinforcement learning.

This is a financial domain language model fine-tuned based on the DeepSeek-R1-Distill-Llama-8B model, optimized using LoRA technology, suitable for financial Q&A and instruction tasks.

Vision instruction fine-tuned model optimized with Unsloth, supporting multimodal task processing

SimPO is a preference optimization method that eliminates the need for reference reward models, simplifying the traditional RLHF pipeline by directly optimizing language models with preference data.

Mistral-SUPRA is a linear RNN model initialized based on Mistral-7B, combining the functions of Transformer and recurrent models.

MoE-LLaVA is a large vision-language model based on the Mixture of Experts architecture, achieving efficient multimodal learning through sparse activation parameters

A few-shot learning text classification model based on the SetFit framework, achieving efficient classification through contrastive learning and classification head training

Godot-Dodo is an instruction-following model fine-tuned from LLaMA 13B, specializing in code instruction understanding and generation tasks

An image classification model provided by Keras, supporting multiple pre-trained architectures and suitable for common image classification tasks.

This is a reinforcement learning model based on the PPO algorithm, designed to solve control problems in the Pendulum-v1 environment.

A dual-encoder dot-product scoring architecture based on DistilBert, trained on the MSMARCO-Passage dataset with balanced topic-aware sampling, suitable for dense retrieval and candidate set re-ranking

DeiT is a data-efficient image Transformer model trained with attention mechanisms, pretrained and fine-tuned on the ImageNet-1k dataset at 224x224 resolution.

This model is based on the BERT-mini architecture, fine-tuned on the SQuAD 2.0 dataset using the M-FAC second-order optimizer for question answering tasks.