Can LLMs Generate Random Numbers? Evaluating LLM Sampling in Controlled Domains

Citation

@inproceedings{llm_sampling_renda_hopkins_2023,
title={Can {LLM}s Generate Random Numbers? Evaluating {LLM} Sampling in Controlled Domains},
author={Renda, Alex and Hopkins, Aspen K. and Carbin, Michael},
booktitle={ICML 2023 Workshop: Sampling and Optimization in Discrete Space},
year={2023},
url={http://people.csail.mit.edu/renda/llm-sampling-paper},
}

Contents of this repository

This repository contains the bulk of the code to reproduce the data (./main.py) and figures (./plot.py) in the paper. It does not contain any of the models evaluated.

The experiments in the paper used llama.cpp at commit 7f0e9a77 and llama-cpp-python at commit a1b2d5c0. However, we have added (relatively untested) support for Huggingface models as well.

Setup

Install dependencies

pip install -r requirements.txt

If you're using llama.cpp models, you'll also need to install llama-cpp-python and have it available in your Python environment.

Download models

Download the models you want to evaluate and place them in the models directory. The models evaluated in the paper are the 8-bit quantized versions of the llama.cpp models, stored in models/llama-7B/ggml-model-q8_0.bin, models/alpaca-7B/ggml-model-q8_0.bin, models/llama-13B/ggml-model-q8_0.bin, etc.

Running the Code

Kick-the-tires

The simplest experiment is to run bert-tiny-uncased on the uniform bits domain:

./main.py --model bert-tiny-uncased --experiment bits_uniform --prompt-examples 0 1 2 3 4 5 6 7 8 9 10 --trials 10 autoregressive --samples 10 --rollouts 10
./main.py --model bert-tiny-uncased --experiment bits_uniform --prompt-examples 0 1 2 3 4 5 6 7 8 9 10 --trials 10 oneshot
./plot.py --domains bits_uniform --show --models bert-tiny-uncased

Generating Data

To evaluate each different sampling technique, run:

./main.py --model {MODEL_NAME} --experiment {EXPERIMENT NAME} --prompt-examples {LIST OF PROMPT EXAMPLES TO EVALUATE WITH} --trials {NUMBER OF TRIALS TO RUN} (oneshot | autoregressive --rollouts {NUMBER OF ROLLOUTS} --samples {NUMBER OF SAMPLES PER ROLLOUTS} )

Where:

• MODEL NAME is a model name listed here
• EXPERIMENT NAME is an experiment listed here
• PROMPT EXAMPLES is a list of numbers (e.g., 0, 0 1, 0 1 2 3 4 5 6 7 8 9 10), which specifies how many prompt examples to evaluate with (providing multiple numbers runs multiple distinct experiments)
• TRIALS is the number of trials to run per configuration
• oneshot says to run NARS experiments
• autoregressive says to run ARS experiments
  • ROLLOUTS is the number of rollouts to run in a given autoregressive experiment
  • SAMPLES is the number of samples to generate per rollout

Plotting Data

To plot the data, run ./plot.py --domains {DOMAINS} [--models {MODELS}] [--show]

Manually Inspecting Data

Data is saved in the data/ directory.

For one-shot (NARS) data:

with open('data/bits_uniform/oneshot/prompt-0/bert-tiny-uncased/trial-0/pcfg-logprobs.pkl', 'rb') as f:
    data = pickle.load(f)
print(data)

Results in:

([(['0'], -0.6931471805599453), (['1'], -0.6931471805599453)],
 [-7.413161754608154, -6.99363374710083])

where the first element of the tuple is the ground-truth log-probabilities for each element of the domain, and the second element is the predicted log-probabilities for those elements.

For autoregressive (ARS) data:

with open('data/bits_uniform/oneshot/prompt-0/bert-tiny-uncased/trial-0/rollouts.pkl', 'rb') as f:
    data = pickle.load(f)
print(data)

will print a list of lists. Each list represents the samples in a given rollout.

Editing the Code

You'll likely have to edit the code to do anything other than replicate the entirety of the paper experiments.

Adding new models

You can add new models by editing here.

Adding new domains

You can add new domains by adding a new entry here. In the current framework, each domain is represented by a PCFG in util.py.