Reasoning-PoT

Program-of-Thought (PoT) fine-tuning and evaluation for GSM8K-style perturbed math data.

Quickstart

Install dependencies:

pip install -r requirements.txt

If needed:

pip install tf-keras

Fine-tune

python finetune_pot.py \
  --model_name meta-llama/Llama-3.1-8B-Instruct \
  --data_path training_data/gsm8k_concrete_training_data.json \
  --output_dir pot-finetuned

Evaluate

python evaluate_prompts.py \
  --model_path pot-finetuned \
  --dataset_path gsm_perturbed_with_new_questions.json \
  --prompt pot

Organization

• finetune_pot.py: training

  • parser commands:
    • --model_name
    • --data_path
    • --prompt_key
    • --completion_key
    • --output_dir
    • --num_epochs
    • --batch_size
    • --learning_rate
    • --lora_r
    • --lora_alpha
    • --lora_dropout
    • --use_4bit
    • --gradient_accumulation_steps
    • --max_length
    • --train_size
    • --eval_size
    • --warmup_steps
    • --logging_steps
    • --save_steps
    • --wandb_project
    • --wandb_run_name
    • --no_wandb

• evaluate_prompts.py: evaluation

  • parser commands:
    • --model_path
    • --dataset_path
    • --prompt
    • --prompt_templates
    • --output_file
    • --device
    • --tolerance
    • --use_base_model
    • --num_attempts
    • --temperature
    • --top_p
    • --max_new_tokens

• prompt_templates.json: prompt templates

• training_data/: example datasets

Data note

This repo does not include the ReasonAgain dataset. You can find it in the CogComp reasoning-eval repository.