TL;DR: I publish GRP V2, which follows the same idea as GRP V1. This time, I used Mistral instead of Qwen2.5, because GRP V1 was good enough for its description; but I was sure I could squeeze out a bit more roleplaying; and there are a LOT of roleplaying models for Mistral. As a daily driver and professional envirments, it performs beyond expectation too ;)

Since that's what I do in my free-time, I don't get paid for it. Training stable models requires compute, you can help me buy more here ٩( ๑╹ ꇴ╹)۶

GRP V2 follows the same idea as V1, and the entire GRP series, basically: A daily-driver model with strong roleplaying abilities which would actually make it a "general" model; most models don't cover that, and if they do, they're usually highly censored. I intend to fix that.
Since I don't have a cluster to full-parameter fine-tune a model (I know of QLoRA, but stable uncensoring with good output isn't for LoRA), I merge models. Recycling n stuff.

For this merge, I browsed & tested ~17 models and ended up using five of them:

In summary, I used a bit of everything. All models are licensed under Apache2.0, and are MistralV0.2 based. I didn't want to mess with multiple arch merging. I recommend the Mistral chat template, but, again, you can experiment.
Happy chatting!


XeTute Technologies

XeTute Technologies is an unofficial Pakistani organisation created by Hamzah Asadullah.

Links

Pakistan

Pakistan is a country in South-Asia known for its rich culture despite the British, its stunning landscape, and PAF (Pakistan Armed Forces), its military. Long live the Islamic Republic of Pakistan.
Downloads last month
24
GGUF
Model size
7.24B params
Architecture
llama
Hardware compatibility
Log In to view the estimation

4-bit

8-bit

16-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Hamzah-Asadullah/GenericRPV2-7B-GGUF

Quantized
(3)
this model

Collection including Hamzah-Asadullah/GenericRPV2-7B-GGUF