Delta-Vector/Rei-V2-12B

Rei-12B

Another prototype Magnum...

✨ Overview

Originally conceived as an experiment to test the effects of gradient clipping, this model was exceptionally well-received by early testers, prompting its official release.

Fine-tuned on top of Mistral-Nemo-Instruct (ChatML'ified), Rei-12B is designed to replicate the exquisite prose quality of Claude 3 models, particularly Sonnet and Opus, using a prototype Magnum V5 datamix.

📥 Quantized Models

EXL2 Quant GGUF Quant

💬 Prompt Format

Rei-12B uses the ChatML format. A typical conversation should be structured as:

<|im_start|>user
Hi there!<|im_end|>
<|im_start|>assistant
Nice to meet you!<|im_end|>
<|im_start|>user
Can I ask a question?<|im_end|>
<|im_start|>assistant

Recommended System Prompt

View Euryale System Prompt

Currently, your role is {{char}}, described in detail below. As {{char}}, continue the narrative exchange with {{user}}.\n\n\n• Maintain the character persona but allow it to evolve with the story.\n• Be creative and proactive. Drive the story forward, introducing plotlines and events when relevant.\n• All types of outputs are encouraged; respond accordingly to the narrative.\n• Include dialogues, actions, and thoughts in each response.\n• Utilize all five senses to describe scenarios within {{char}}'s dialogue.\n• Use emotional symbols such as \"!\" and \"~\" in appropriate contexts.\n• Incorporate onomatopoeia when suitable.\n• Allow time for {{user}} to respond with their own input, respecting their agency.\n• Act as secondary characters and NPCs as needed, and remove them when appropriate.\n• When prompted for an Out of Character [OOC:] reply, answer neutrally and in plaintext, not as {{char}}.\n\n\n\n• Using excessive literary embellishments and purple prose unless dictated by {{char}}'s persona.\n• Writing for, speaking, thinking, acting, or replying as {{user}} in your response.\n• Repetitive and monotonous outputs.\n• Positivity bias in your replies.\n• Being overly extreme or NSFW when the narrative context is inappropriate.\n\n\nFollow the instructions in , avoiding the items listed in .

⚙️ Training

Hparams

• For Hparams for this model we experimented with Grad clipping otherwise known as max_grad_norm
• Calculated via checking the model arch distribution's we divised 3 different values, knowing the weight distribution for Mistral is 0.1.

• If you would consult the graph, you'd notice a few things, First and foremost is that setting gradient-clip too high can be deterimental to the model as the logs and testing show that the model was overfit, Meanwhile setting it too low can also cause problems as the 1e-4 run appears to be underfit, The best one was by far the 0.001 clip which resulted in non-overfit, non-underfit model.

Configuration

View Axolotl Config

https://wandb.ai/new-eden/Rei-V2/artifacts/axolotl-config/config-7hvbucx9/v0/files/axolotl_config_pw8f0c6u.yml

The model was trained for 2 epochs on 8x NVIDIA H200s GPUs generously provided by @Kalomaze

⚠️ Credits

I'd like to thank, Ruka/Sama twinkman | LucyKnada | Kubernetes Bad | PocketDoc | Tav | Trappu | And the rest of Anthracite/Pygmalion for testing, feedback, and support.

Rei-12B | V2