ruggsea/Llama3.1-Instruct-SEP-Chat

Llama3-SEP-Chat: Philosophy Expert Assistant

This model is a LoRA-finetuned version of meta-llama/Meta-Llama-3.1-8B-instruct on a curated dataset of Stanford Encyclopedia of Philosophy (SEP) conversations. The model is designed to engage in philosophical discussions with a formal yet accessible tone, leveraging the comprehensive knowledge from SEP.

Model Description

The model was trained using direct finetuning on the instruct variant of Llama 3, preserving its native chat format and instruction-following capabilities while enhancing its philosophical expertise.

Training Dataset

The training data consists of multi-turn conversations derived from the Stanford Encyclopedia of Philosophy, formatted as chat interactions between a user and an assistant. The conversations maintain academic rigor while ensuring accessibility.

Chat Format

The model uses Llama 3's native chat format, which is automatically applied by the tokenizer. No additional tokens or formatting were added during finetuning.

Training Details

Model Configuration

• Base Model: meta-llama/Meta-Llama-3.1-8B-instruct
• Training Type: LoRA (Low-Rank Adaptation)
• Quantization: 4-bit (NF4)
• Compute: Mixed Precision (bfloat16)

Training Hyperparameters

• Learning Rate: 2e-5
• Train Batch Size: 16
• Gradient Accumulation Steps: 2
• Effective Batch Size: 32
• Optimizer: paged_adamw_8bit
• Training Epochs: 5
• Warmup Ratio: 0.03
• LoRA Configuration:
  • Rank: 256
  • Alpha: 128
  • Dropout: 0.05
  • Target: all-linear layers

Framework Versions

• Transformers: latest
• PEFT: latest
• PyTorch: 2.1.0+cu121
• TRL: latest
• Accelerate: latest

Usage

The model can be used with the standard Hugging Face transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "ruggsea/Llama3.1-Instruct-SEP-Chat"

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Format your input using the chat template
messages = [
    {"role": "user", "content": "What is the categorical imperative?"}
]

# Apply the chat template
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False
)

# Generate response
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
    repetition_penalty=1.2,
    no_repeat_ngram_size=3,
)

response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)

Limitations

• The model's knowledge is primarily focused on philosophical concepts and may not perform as well on general topics
• As with all language models, it may occasionally generate incorrect or inconsistent information
• The model inherits any limitations and biases present in the base Llama 3 model and the SEP dataset

License

This model is subject to the Meta Llama 3 license terms. Please refer to Meta's licensing for usage requirements and restrictions.