ruggsea/Llama3.1-8B-SEP-Chat

Llama3.1-SEP-Chat

This model is a LoRA finetune of meta-llama/Meta-Llama-3.1-8B trained on multi-turn philosophical conversations. It is designed to engage in philosophical discussions in a conversational yet rigorous manner, maintaining academic standards while being accessible.

Model description

The model was trained using the TRL (Transformer Reinforcement Learning) library's chat template, enabling it to handle multi-turn conversations in a natural way. It builds upon the capabilities of its predecessor Llama3-stanford-encyclopedia-philosophy-QA but extends it to handle more interactive, back-and-forth philosophical discussions.

Chat Format

The model uses the standard chat format with roles:

<|system|>
{{system_prompt}}
<|user|>
{{user_message}}
<|assistant|>
{{assistant_response}}

Training Details

The model was trained with the following system prompt:

You are an expert and informative yet accessible Philosophy university professor. Students will engage with you in philosophical discussions. Respond to their questions and comments in a correct and rigorous but accessible way, maintaining academic standards while fostering understanding.

Training hyperparameters

The following hyperparameters were used during training:

• —Learning rate: 2e-5
• —Train batch size: 1
• —Gradient accumulation steps: 4
• —Effective batch size: 4
• —Optimizer: pagedadamw8bit
• —LR scheduler: cosine with warmup
• —Warmup ratio: 0.03
• —Training epochs: 5
• —LoRA config:
• —r: 256
• —alpha: 128
• —Target modules: all-linear
• —Dropout: 0.05

Framework versions

• —PEFT 0.10.0
• —Transformers 4.40.1
• —PyTorch 2.2.2+cu121
• —TRL latest
• —Datasets 2.19.0
• —Tokenizers 0.19.1

Intended Use

This model is designed for:

• —Multi-turn philosophical discussions
• —Academic philosophical inquiry
• —Teaching and learning philosophy
• —Exploring philosophical concepts through dialogue

Limitations

• —The model should not be used as a substitute for professional philosophical advice or formal philosophical education
• —While the model aims to be accurate, its responses should be verified against authoritative sources
• —The model may occasionally generate plausible-sounding but incorrect philosophical arguments
• —As with all language models, it may exhibit biases present in its training data

License

This model is subject to the Meta Llama 2 license agreement. Please refer to Meta's licensing terms for usage requirements and restrictions.

How to use

Here's an example of how to use the model:

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model = AutoModelForCausalLM.from_pretrained("ruggsea/Llama3.1-SEP-Chat")
tokenizer = AutoTokenizer.from_pretrained("ruggsea/Llama3.1-SEP-Chat")


# Example conversation
messages = [
    {"role": "user", "content": "What is the difference between ethics and morality?"}
]

# Format prompt using chat template
prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False
)

# Generate response
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)