Models by this creator




Total Score


Llama-3-8b-64k-PoSE is a large language model (LLM) developed by winglian that extends the context length of the Llama 3 8B model from 8k to 64k tokens using Packed Sparse Attention (PoSE). The model was trained on a subset of the RedPajama v1 dataset with text between 6k-8k tokens, and further fine-tuned with a rank stabilized LoRA. Compared to the base Llama 3 8B model, this extended context version can handle longer input sequences. Similar models include the Meta-Llama-3-8B and Meta-Llama-3-70B models, which are also part of the Llama 3 family developed by Meta. These models come in 8B and 70B parameter sizes and have both pre-trained and instruction-tuned versions. Model inputs and outputs Inputs The model takes in text input only. Outputs The model generates text and code. Capabilities Llama-3-8b-64k-PoSE can handle longer input sequences than the base Llama 3 8B model due to its extended 64k token context length. This makes it well-suited for tasks that require processing of long-form text, such as summarization, question answering on lengthy passages, or text generation with large context windows. What can I use it for? The extended context capabilities of Llama-3-8b-64k-PoSE make it a good choice for applications that need to work with long-form text, such as academic writing assistance, long-form journalism, or analysis of lengthy documents. Developers could fine-tune the model further for specific use cases to leverage its ability to maintain coherence and context over longer spans of text. Things to try One interesting aspect of this model is the use of Packed Sparse Attention (PoSE) to extend the context length. Developers could experiment with different PoSE hyperparameters or explore other techniques for increasing the context window of large language models. Additionally, the model's performance on tasks that require long-range understanding, such as multi-document summarization or long-form question answering, would be an interesting area to investigate further.

Read more

Updated 5/27/2024