# Instruction Fine-tuning Optimization

A 0.5B parameter-scale text generation model based on llama.cpp, supporting instruction-based text generation tasks

Llama-3.1-8B model fine-tuned using the automatically filtered 50,000-entry Tulu-3-MIG dataset

Captain Violet is a 12B-parameter merged model, created by combining Epiculous/Violet_Twilight-v0.2 and Nitral-AI/Captain_BMO-12B using the mergekit tool, supporting text generation tasks.

The Ichigo-llama3s series model is a multimodal language model developed by Homebrew Research, natively supporting audio and text input comprehension. Based on the Llama-3 architecture, it is trained using WhisperVQ as an audio file tokenizer, enhancing its audio understanding capabilities.

Tarsier-34b is an open-source large-scale video-language model focused on generating high-quality video captions and achieving leading results in multiple public benchmarks.

Mistral-7B-Instruct-v0.3 is an instruction-tuned version of Mistral-7B, focusing on text generation tasks that follow human instructions.

A large-scale language model developed by the Japanese collaborative project LLM-jp, supporting Japanese and English, primarily used for text generation tasks.

Wizard Lake 7B is a fusion model combining the next-generation WizardLM 2 7B model with a customized DolphinLake model, delivering outstanding performance.

Idefics2 is an open-source multimodal model capable of accepting arbitrary sequences of image and text inputs to generate text outputs. It shows significant improvements in OCR, document understanding, and visual reasoning.

This model is a GGUF format model converted from Open-Orca/Mistral-7B-OpenOrca, suitable for text generation tasks.

Code Llama is a series of code generation and understanding models developed by Meta, ranging from 7 billion to 34 billion parameters. This model is the 34 billion parameter instruction fine-tuned version.

A language model fine-tuned based on the Mixture of Experts (MoE) architecture, which fuses multiple models through the LazyMergekit framework and performs excellently in instruction tasks.

Mistral-7B-Instruct-v0.2 is an enhanced instruction fine-tuned large language model based on Mistral-7B-Instruct-v0.1, achieving 30% sparsity through Wanda pruning method without requiring retraining while maintaining competitive performance.

MindLLM 1.3B is a 1.3 billion-parameter Transformer model jointly developed by the Beijing Engineering Research Center of Massive Language Information Processing and Cloud Computing Applications and the Southeast Institute of Information Technology, Beijing Institute of Technology, supporting Chinese and English dialogue generation.

Phind-CodeLlama-34B-v2 is an open-source code generation model fine-tuned from Phind-CodeLlama-34B-v1, achieving 73.8% pass@1 on HumanEval tests, representing the state-of-the-art among open-source models.

An open-source long-context language model fine-tuned based on Meta's original Llama-2 7B model, supporting 32K context length

Fully open-source and commercially usable Chinese version of the Llama2 model with Chinese-English supervised fine-tuning dataset. The input format strictly adheres to the llama-2-chat standard and is fully compatible with all optimization solutions for the original llama-2-chat model.

CodeT5+ is an open-source family of large language models for code, supporting code understanding and generation tasks, featuring an encoder-decoder architecture with flexible mode switching.

A 7-billion-parameter open-source dialogue model fine-tuned from EleutherAI Pythia-7B, trained on over 40 million instructions using 100% carbon-negative computing resources