llm_reasoning

A repo to understand the reasoning capabilities of llms

Select and Deduct

The implemented method for this section is explained in this blog post. The general idea is to implement a simple, modular and explainable reasoning system that tries to mimic the reasoning capabilities of humans in a combination of forward and backward reasoning.

Training the models yourself

Downloading Data

To prepare the dataset for training the models, download both ProofWriter and EntailmentBank datasets from this link and extract both of the folders inside the data folder. The ProofWriter dataset is post-processed according to the NLProofs repo to match the same format as the EntailmentBank dataset.

Preprocessing and Training

The select_and_deduct/prepare_data_for_seq2seq_training.sh script will preprocess the downloaded data to prepare augmented datasets containing a balanced combination of both ProofWriter and EntailmentBank datasets from all subtasks. For ProofWriter we take all correct proofs from all depths and sample it with the same number of examples as the EntailmentBank dataset. For the EntailmentBank dataset, we take all the examples from subtask 1 and 2.

The script will use select_and_deduct/deductor_data_handler.py and select_and_deduct/selector_data_handler.py to prepare the data. By default, the deductor will append the final hypothesis to the input which you can control by --add_hypothesis flag. After preparing data, you will get 4 files in data/ directory which are training and validation files for both the selector and the deductor.

To train the models, you can use the select_and_deduct/run_train_deductor_flant5.sh and select_and_deduct/run_train_selector_flant5.sh scripts which will train a flan-t5 model for both tasks.

Inference

You can run the proof search algorithm by running the select_and_deduct/proof_search.py script. Here are a few important arguments to control the search algorithm:

PYTHONPATH=.. python proof_search.py \
  --deductor_path ../outputs/deductor-flant5-xl/ \
  --selector_path ../outputs/selector-flant5-xl/ \
  --test_data_path ../data/test/ent_task2.json \
  --output_dir ent_task2_hyp_th0.5_beamsearch/ \
  --n_search_beams 8 \
  --hypothesis_acceptance_threshold 0.5 \
  --search_algorithm beam_search \
  --deductor_add_hyp

You can add or remove hypothesis from the input of the deductor by using the --deductor_add_hyp flag with respect to the training configuration. Also, you can control the number of search beams by using the --n_search_beams flag. The --search_algorithm flag can be set to beam_search, greedy or early_selection. You can also set the deductor and selector paths to the trained models or the pretrained models we discuss in the next section.

Using the pretrained models

Both selector and deductor models are trained and shared on huggingface so that you can use them directly (deductor model, selector model). You can use the code inside select_and_deduct/inference.py to run a gradio server and interact with the models with your custom inputs. It will load the pretrained models and run the proof search algorithm to find the proofs for the given input and plots the entailment graph along with intermediate steps as shown below: