Iridium-72B-v0.1

Model Description

Iridium is a 72B parameter language model created through a merge of Qwen2-72B-Instruct, calme2.1-72b, and magnum-72b-v1 using model_stock.

Features

• 72 billion parameters
• Combines Magnum prose with Calam smarts

Technical Specifications

Architecture

• Qwen2ForCasualLM
• Models: Qwen2-72B-Instruct (base), calme2.1-72b, magnum-72b-v1
• Merged layers: 80
• Total tensors: 963
• Context length: 128k

Tensor Distribution

• Attention layers: 560 files
• MLP layers: 240 files
• Layer norms: 160 files
• Miscellaneous (embeddings, output): 3 files

Merging

Custom script utilizing safetensors library.

Usage

Loading the Model

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained("leafspark/Iridium-72B-v0.1", 
                                             device_map="auto", 
                                             torch_dtype=torch.float16)
tokenizer = AutoTokenizer.from_pretrained("leafspark/Iridium-72B-v0.1")

GGUFs

Find them here: leafspark/Iridium-72B-v0.1-GGUF

Optimal Sampling Parameters

I found these to work well:

{
  "temperature": 1
  "min_p": 0.08
  "top_p": 1
  "top_k": 40
  "repetition_penalty": 1
}

Hardware Requirements

• At least 135GB of free space
• ~140GB VRAM/RAM