Covenant-72B-Chat is the instruction-tuned variant of
Covenant-72B, the largest
permissionless collaboratively trained language model. It was fine-tuned via
supervised fine-tuning (SFT) on top of the 72B-parameter base model.
For more details, see the
technical report.
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer
3
4model = AutoModelForCausalLM.from_pretrained(
5 "1Covenant/Covenant-72B-Chat",
6 torch_dtype=torch.bfloat16,
7 device_map="auto",
8)
9tokenizer = AutoTokenizer.from_pretrained("1Covenant/Covenant-72B-Chat")
10
11messages = [
12 {"role": "user", "content": "Explain general relativity in simple terms."},
13]
14input_ids = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
15output_ids = model.generate(input_ids, max_new_tokens=256)
16print(tokenizer.decode(output_ids[0][input_ids.shape[-1]:], skip_special_tokens=True))
