# Instruction fine-tuning optimization

Gervásio 8B PTPT is an open-source decoder model for Portuguese, fine-tuned based on LLaMA 3.1 8B Instruct, with powerful text generation capabilities.

Falcon-H1-34B-Instruct is a large language model with 34B parameters, focusing on instruction-following tasks.

This is a 12B-parameter large language model, processed with llama.cpp's imatrix quantization, offering multiple quantized versions to accommodate different hardware requirements.

This is a merged version based on multiple 12B-parameter scale models, integrating 7 different pre-trained language models with distinct characteristics using the model_stock method to enhance comprehensive capabilities.

Optimized Qwen2 model based on Unsloth and Huggingface TRL library, achieving 2x inference speed improvement

QwQ-32B is a 32B-parameter large language model based on the Qwen2 architecture, processed with 4-bit integer quantization using the GPTQ method, suitable for efficient text generation tasks.

The Nemotron-UltraLong-8B series is a language model specifically designed for processing ultra-long text sequences, supporting a context window of up to 4 million tokens while maintaining exceptional performance.

This is a merged model based on Mistral-Small-24B-Instruct-2501, utilizing the DARE TIES fusion method to integrate multiple 24B parameter-scale models.

A Greek instruction-tuned large language model based on Llama-3.1-8B, enhancing Greek language capabilities and supporting multilingual tasks

This project aims to create a language model with better performance by fusing multiple pre-trained language models of 70B scale. Based on the Llama 3.3 instruction model, the Linear DELLA fusion method is used for model fusion.

SEA-LION is a series of large language models pre-trained and instruction fine-tuned for the Southeast Asian region, dedicated to solving multilingual processing problems in this region and providing strong support for natural language processing of Southeast Asian languages.

A multimodal language model based on Llama-3 architecture, supporting audio and text input understanding with noise robustness and multi-turn dialogue capabilities

Bllossom-3B is a Korean-English bilingual enhanced version based on meta-llama/Meta-Llama-3.2-3B, trained through full parameter fine-tuning and curated Korean data, fully retaining English capabilities while enhancing Korean language processing.

The Ichigo-llama3s series models natively support audio and text input comprehension, based on the Llama-3 architecture, using WhisperVQ as the tokenizer for audio files.

This is a statically quantized version based on behbudiy/Llama-3.1-8B-Instuct-Uz, supporting Uzbek and English, suitable for various text generation tasks.

Solar Pro Preview is an advanced large language model with 22 billion parameters, designed specifically for single-GPU operation and delivering excellent performance.

EuroLLM-1.7B is the first pre-trained model in the EuroLLM series, with multilingual processing capabilities, capable of understanding and generating text in multiple European and other related languages.

Mistral-Nemo-Base-2407 is a 12-billion-parameter generative text pre-training model jointly trained by Mistral AI and NVIDIA, outperforming models of similar or smaller scale.

Meta-Llama-3.1-405B-Instruct is a large language model with 405 billion parameters based on the Llama 3.1 architecture, optimized for instruction-following tasks and supporting multiple languages.

Llama-3.1-8B-Instruct is an 8B parameter large language model released by Meta, focusing on instruction-following tasks.

This is a merged pre-trained language model built on Meta Llama 3, which combines the advantages of the SimPO and SPPO-Iter3 models and is suitable for text generation tasks.

UCCIX-Llama2-13B-Instruct is an Irish-English bilingual large language model, developed based on the Llama 2-13B architecture, with special optimizations for Irish language processing.

A Chinese role-playing optimization model fine-tuned on instructions based on hfl/llama-3-chinese-8b-instruct-v3

An 8B parameter model based on the Llama 3 architecture, trained with various DPO datasets, focused on text generation tasks

Ablated version of WizardLM-2-7B, processed with orthogonalization to optimize specific behavioral patterns

XGen-MM is the latest multimodal large model series developed by Salesforce AI Research. Based on the successful design of BLIP, it achieves a more powerful and superior model architecture through fundamental enhancements.

The 8-bit quantized version of Meta Llama 3 - 8B Instruct large language model, reducing model size and improving inference speed, suitable for deployment on resource-constrained devices.

Idefics2 is an open multimodal model capable of accepting arbitrary sequences of images and text as input and generating text output. The model can answer questions about images, describe visual content, create stories based on multiple images, or function purely as a language model.

A merged model based on Llama-3-8b, integrating multiple pre-trained language models using the Model Stock method

Mistral-7B-Instruct-v0.2 model quantized to FP8 precision by FriendliAI, significantly improving inference efficiency while maintaining high accuracy.

This is a hybrid model created by merging multiple Mistral-7B variant models using the mergekit tool, focusing on text generation tasks.

A version fine-tuned with German instructions based on Google's Gemma 2B model, exploring the possibility of German-to-English translation through vocabulary expansion.

Gemma is a lightweight open-source large language model launched by Google. It is built on the same technology as Gemini and is suitable for various text generation tasks.

Jais-30b-chat-v3 is a large language model fine-tuned on selected Arabic and English Q&A datasets based on Jais-30b-v3. It is optimized for Arabic and English conversations and has the ability to process long contexts of 8000 tokens.

Gemma is a lightweight open model series launched by Google, built on the same technology as Gemini, suitable for text generation tasks.

Gemma is a lightweight open-source large language model series launched by Google, built on the technology used to create Gemini models, offering a base version with 2 billion parameters.

MalayaLLM is a language model focused on generative AI for Malayalam, built through continued pre-training on the LLAMA2 model with the addition of approximately 18,000 Malayalam vocabulary entries.

Mistral-7B-Instruct-v0.2 is an instruction-finetuned large language model improved from Mistral-7B-Instruct-v0.1, compressed to 2% sparsity using Wanda pruning method while maintaining competitive performance without retraining.

MobileLLaMA-1.4B-Chat is a chat model fine-tuned from MobileLLaMA-1.4B-Base, utilizing the ShareGPT dataset for supervised instruction fine-tuning.

A 200-million-parameter Chinese causal language model based on the Phi2 architecture, supporting text generation tasks