TowerInstruct-7B-v0.2-GGUF

Model Info

• Model creator: Unbabel
• Original card (has more info): https://huggingface.co/Unbabel/TowerInstruct-7B-v0.2
• Languages: English, Portuguese, Spanish, French, German, Dutch, Italian, Korean, Chinese, Russian.
• Context size: 4096, but according to the model's card it was trained with max_seq_length: 2048, so make sure your input is ≤2048 tokens (is that ~1500 words?).
• Template: The model was trained using the ChatML prompt templates WITHOUT ANY SYSTEM PROMPTS !!!, i.e. there's no <|im_start|>system, it's just

<|im_start|>user
{USER PROMPT}<|im_end|>
<|im_start|>assistant
{MODEL RESPONSE}<|im_end|>
<|im_start|>user
[...]

Quantization info

I didn't use imatrix because I have no idea if they are okay for non-general-purpose LLMs.

Quantized with llama.cpp @ 4c4cb30736582cacb1a164a9d4bc8e17b1014be7 (2024-02-24).

IQ3_M requires minimum 4c4cb30736582cacb1a164a9d4bc8e17b1014be7

Licensing

These .gguf files themselves are under CC-0 (i.e. public domain).

TowerInstruct-7B-v0.2 itself is under CC-BY-NC-4.0 ("NC" in "CC-BY-NC-4.0" means "non-commercial" which is legally ambiguous),

but it is based on Llama 2 model which is licensed under the LLAMA 2 Community License (Copyright © Meta Platforms, Inc. All Rights Reserved.) whose TL;DR is you may not use it for bullying, military, and criminal activity.