yodayo-ai/nephra_v1.0

Overview

nephra v1 is primarily a model built for roleplaying sessions, trained on roleplay and instruction-style datasets.

Model Details

• Developed by: Sao10K
• Model type: Text-based Large Language Model
• License: Meta Llama 3 Community License Agreement
• Finetuned from model: Meta-Llama-3-8B

Inference Guidelines

import transformers
import torch

model_id = "yodayo-ai/nephra_v1.0"

pipeline = transformers.pipeline(
  "text-generation",
  model=model_id,
  model_kwargs={"torch_dtype": torch.bfloat16},
  device_map="auto",
)

messages = [
  {"role": "system", "content": "You are to play the role of a cheerful assistant."},
  {"role": "user", "content": "Hi there, how's your day?"},
]

prompt = pipeline.tokenizer.apply_chat_template(
  messages,
  tokenize=False,
  add_generation_prompt=True
)

outputs = pipeline(
  prompt,
  max_new_tokens=512,
  eos_token_id=[
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>"),
    pipeline.tokenizer.eos_token_id,
  ],
  do_sample=True,
  temperature=1.12,
  min_p=0.075,
)
print(outputs[0]["generated_text"][len(prompt):])

Recommended Settings

To guide the model to generate high-quality responses, here are the ideal settings:

Prompt Format: Same Prompt Format as Llama-3-Instruct
Temperature - 1.12
min-p: 0.075
Repetition Penalty: 1.1
Custom Stopping Strings: "\n{{user}}", "<" , "```" , -> Has occasional broken generations.

License

Nephra v1 falls under META LLAMA 3 COMMUNITY LICENSE AGREEMENT.