Skip to content

Few-shot setting, Chat Templates and Batch evaluation

Latest
Compare
Choose a tag to compare
@ikergarcia1996 ikergarcia1996 released this 02 Mar 11:24

Changelog

  1. Use the tokenizer chat template found in the tokenizer_config.json file instead of FastChat. This change eliminates the need for users to specify the chat template for each model, ensuring the correct template is always used.

  2. Implement the few-shot setting, in which the model is prompted with 4 examples from the training dataset of each split, totaling 44 few-shot examples. The few-shot examples for each split correspond to:

Distractor: False. Negation: False 
Distractor: False. Negation: True 
Distractor: True. Negation: False 
Distractor: True. Negation: True
  1. Implement a method to evaluate multiple models without the need to create a configuration for each one:
for model_name in \
meta-llama/Llama-2-70b-chat-hf \
meta-llama/Llama-2-70b-hf \
meta-llama/Llama-2-13b-chat-hf \
meta-llama/Llama-2-13b-hf \
meta-llama/Llama-2-7b-chat-hf \
meta-llama/Llama-2-7b-hf \
mistralai/Mistral-7B-Instruct-v0.2 \
mistralai/Mixtral-8x7B-Instruct-v0.1 \
do

accelerate launch run.py --config configs/zero-shot/base.yaml --model_name_or_path "$model_name" --output_dir results/zero-shot/"$model_name"

done
  1. Other minor fixes to ensure compatibility with additional models and enhance performance.

Results

Zero Shot

Zero-shot

Few Shot

Few-shot