Calme Rys 78 B Orpo v0.1

Introduction

CalmeRys-78B-Orpo-v0.1 is a fine-tuned model based on MaziyarPanahi's Calme-2.4-Rys-78b, designed for various text generation tasks. It offers capabilities such as agentic support, roleplaying, reasoning, multi-turn conversations, and maintaining long context coherence. It ranks highly on the Open LLM Leaderboard as of October 2024.

Architecture

This model is built using the Transformer's library and uses a dataset named mlabonne/orpo-dpo-mix-40k. It is configured to handle text generation tasks and supports quantization optimizations for efficient deployment.

Training

The model was fine-tuned on 1.5k rows from the mlabonne/orpo-dpo-mix-40k dataset. Training visualizations and experiments are documented on Weights & Biases. The model's performance metrics across various datasets, such as IFEval and MATH Level 5, are listed on the Open LLM Leaderboard.

Guide: Running Locally

1. Install Dependencies:

   pip install -qU transformers accelerate bitsandbytes
   huggingface-cli download dfurman/CalmeRys-78B-Orpo-v0.1
   

2. Setup the Model:

   from transformers import AutoTokenizer, BitsAndBytesConfig
   import torch
   import transformers
   
   if torch.cuda.get_device_capability()[0] >= 8:
       !pip install -qqq flash-attn
       attn_implementation = "flash_attention_2"
       torch_dtype = torch.bfloat16
   else:
       attn_implementation = "eager"
       torch_dtype = torch.float16
   
   model = "dfurman/CalmeRys-78B-Orpo-v0.1"
   tokenizer = AutoTokenizer.from_pretrained(model)
   pipeline = transformers.pipeline(
       "text-generation",
       model=model,
       model_kwargs={
           "torch_dtype": torch_dtype,
           "device_map": "auto",
           "attn_implementation": attn_implementation,
       }
   )
   

3. Example Usage: Use the pipeline to generate text based on input prompts. Examples provided include numerical comparisons, table completions, and recipe suggestions.

4. Suggested Cloud GPUs: Consider using cloud GPU services such as AWS EC2 with Tesla V100 or A100 instances for optimal performance.

License

The model is licensed under the MIT License, allowing for broad usage and modification rights.