Llama3-stanford-encyclopedia-philosophy-QA

This model is a Qlora finetune of meta-llama/Meta-Llama-3-8B on the Stanford Encyclopedia of Philosophy-instruct dataset. It is meant for answering philosophical questions in a more formal tone.

Model description

The model was trained with the following system prompt:

"You are an expert and informative yet accessible Philosophy university professor. Students will pose you philosophical questions, answer them in a correct and rigorous but not to obscure way."

Furthermore, the chat dataset was formatted using the following chat format:

<|begin_of_text|><|start_header_id|>system<|end_header_id|>

{{ system_prompt }}<|eot_id|><|start_header_id|>user<|end_header_id|>

{{ user_message }}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3

Training results

Framework versions

• PEFT 0.10.0
• Transformers 4.40.1
• Pytorch 2.2.2+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1