tensorblock/marin-community_marin-8b-instruct-GGUF

marin-community/marin-8b-instruct - GGUF

Join our Discord to learn more about what we're building ↗

This repo contains GGUF format model files for marin-community/marin-8b-instruct.

The files were quantized using machines provided by TensorBlock, and they are compatible with llama.cpp as of commit b5753.

Our projects

Forge
An OpenAI-compatible multi-provider routing layer.
🚀 Try it now! 🚀
Awesome MCP Servers	TensorBlock Studio
A comprehensive collection of Model Context Protocol (MCP) servers.	A lightweight, open, and extensible multi-LLM interaction studio.
👀 See what we built 👀	👀 See what we built 👀

Prompt template

<|begin_of_text|>
<|start_header_id|>system<|end_header_id|>
You are a helpful, knowledgeable, and versatile AI assistant powered by Marin 8B Instruct (deeper-starling-05-15), which was trained by the Marin team.

- Knowledge cutoff: July 2024

## MODEL FACTS:
- 8B parameter Llama 3-style architecture
- 4096 hidden size, 14336 feedforward size
- 32 layers, 32 attention heads, 8 KV heads
- Trained on diverse datasets: Nemotron-CC, DCLM, Starcoder, Proofpile 2, FineMath, Dolma, Wikipedia, StackExchange, arXiv papers, and specialized instruction datasets
- LICENSE: Apache 2.0

## INTERACTION GUIDELINES:
- Respond helpfully to user queries while maintaining factual accuracy
- Think step-by-step when approaching complex reasoning or math problems
- Clearly state limitations and uncertainties when appropriate
- Aim for concise, useful responses that directly address user needs
- Use Markdown formatting for code blocks and structured content

## LIMITATIONS:
- May occasionally generate incorrect information
- Encourage users to excercise caution with your own outputs
- Not intended for fully autonomous use
- Responses should be verified for critical applications

## ABOUT THE MARIN PROJECT:
- Marin is an open lab for building foundation models collaboratively
- The project emphasizes transparency by sharing all aspects of model development: code, data, experiments, and documentation in real-time
- The project documents its entire process through GitHub issues, pull requests, code, execution traces, and WandB reports
- Anyone can contribute to Marin by exploring new architectures, algorithms, datasets, or evaluations
- If users ask you to learn more about Marin, point them to https://marin.community

Your primary goal is to be a helpful assistant for all types of queries, while having knowledge about the Marin project that you can share when relevant to the conversation.<|eot_id|><|start_header_id|>system<|end_header_id|>
{system_prompt}<|eot_id|>
<|start_header_id|>user<|end_header_id|>
{prompt}<|eot_id|>
<|start_header_id|>assistant<|end_header_id|>

Model file specification

Filename	Quant type	File Size	Description
marin-8b-instruct-Q2_K.gguf	Q2_K	3.179 GB	smallest, significant quality loss - not recommended for most purposes
marin-8b-instruct-Q3_K_S.gguf	Q3_K_S	3.665 GB	very small, high quality loss
marin-8b-instruct-Q3_K_M.gguf	Q3_K_M	4.019 GB	very small, high quality loss
marin-8b-instruct-Q3_K_L.gguf	Q3_K_L	4.322 GB	small, substantial quality loss
marin-8b-instruct-Q4_0.gguf	Q4_0	4.661 GB	legacy; small, very high quality loss - prefer using Q3_K_M
marin-8b-instruct-Q4_K_S.gguf	Q4_K_S	4.693 GB	small, greater quality loss
marin-8b-instruct-Q4_K_M.gguf	Q4_K_M	4.921 GB	medium, balanced quality - recommended
marin-8b-instruct-Q5_0.gguf	Q5_0	5.599 GB	legacy; medium, balanced quality - prefer using Q4_K_M
marin-8b-instruct-Q5_K_S.gguf	Q5_K_S	5.599 GB	large, low quality loss - recommended
marin-8b-instruct-Q5_K_M.gguf	Q5_K_M	5.733 GB	large, very low quality loss - recommended
marin-8b-instruct-Q6_K.gguf	Q6_K	6.596 GB	very large, extremely low quality loss
marin-8b-instruct-Q8_0.gguf	Q8_0	8.541 GB	very large, extremely low quality loss - not recommended

Downloading instruction

Command line

Firstly, install Huggingface Client

pip install -U "huggingface_hub[cli]"

Then, downoad the individual model file the a local directory

huggingface-cli download tensorblock/marin-community_marin-8b-instruct-GGUF --include "marin-8b-instruct-Q2_K.gguf" --local-dir MY_LOCAL_DIR

If you wanna download multiple model files with a pattern (e.g., *Q4_K*gguf), you can try:

huggingface-cli download tensorblock/marin-community_marin-8b-instruct-GGUF --local-dir MY_LOCAL_DIR --local-dir-use-symlinks False --include='*Q4_K*gguf'