Skip to content

Latest commit

 

History

History
163 lines (103 loc) · 5.35 KB

README.md

File metadata and controls

163 lines (103 loc) · 5.35 KB
Raw

Diachronic Testing of Causal Language Models

This repository is a fork from Prompting is not a substitute for probability measurements in large language models by Hu & Levy (2023).

Please refer to the author's original implementation for details.

This work in progress fork adds backends for Pythia and OLMo (with option for running quantized) models via the transformers library. It is primarily meant for prototyping evaluations using the existing preprocessed datasets from Hu & Levy.

Overview

Additional Models

EleutherAI-Pythia

AI2-OLMo

Both models were released in different parameter sizes and with intermediate checkpoints available. The weights, training code, and data are all fully accessible.

Setup

  • requires GPU with cuda >= 12.1 support (smaller models can run on CPU, but not recommended)

venv

uv venv
# macOS / Linux
source .venv/bin/activate
# Windows
.venv\Scripts\activate
uv pip install -r requirements.txt

conda

conda env create -f environment.yml
conda activate metalinguistic-prompting

Evaluation materials

From the original authors:

Evaluation datasets can be found in the datasets folder. Please refer to the README in that folder for more details on how the stimuli were assembled and formatted.

Evaluation scripts

The scripts folder contains scripts for running the experiments. Results of Experiments 1, 2, and 3b can be visualized with new_analysis.ipynb, but visualization of Experiment 3a (isolated instances) is currently not supported for the new models (maybe tbd).

Experiments

  • Template

    bash scripts/run_exp{1,2,3a,3b}_hf.sh {corpus} {huggingface/model}
# optional: checkpoint {revision}, quantization {4bit, 8bit}

  • Example calls

    • Experiment 1

      {corpus}: {p18, news}

      bash scripts/run_exp1_hf.sh news EleutherAI/pythia-70m-deduped step3000

    • Experiment 2

      bash scripts/run_exp2_hf.sh google/flan-t5-small

    • Experiment 3

      {corpus}: {syntaxgym, blimp}

      bash scripts/run_exp3a_hf.sh syntaxgym allenai/OLMo-7B-hf main 8bit

Pythia & OLMo models

  • revision

    • main corresponds to the final model checkpoint. Must be set when using quantization. Check either Pythia or OLMo model cards on Huggingface for details on how to access different (earlier) checkpoints.

  • quantization

    • 8bit or 4bit, running with less precision also requires less VRAM. Loading checkpoint shards can take longer than with full precision (quantized OLMo models load fine, Pythia models very slow). Must set revision to use.

OpenAI

probably needs fixing

The original OpenAI implementation (*_openai.sh) used 3 different models (text-curie-001/GPT-3, text-davinci-002/GTP-3.5-SFT, text-davinci-003/GTP-3.5-SFT+RLHF) all of which are deprecated by now.

There are still 2 base models available via OpenAI's API (babbage-002/replacement for GPT-3 ada and babbage base models and davinci-002/replacement for GPT-3 curie and davinci base models), however, the scripts need to be updated first (not done yet). You also need to provide an API key. See the original repository for details.

For more details on the base models still available read the official documentation.

ToDo

  • test minicons implementation

  • test instruct-tuned models for all other prompting techniques than direct

  • fix Pythia quantization - loading quantized checkpoint shards for Pythia takes too long (works for OLMo though)

  • add batching support - only single instances passed to the model, possible improvements achievable (especially for larger models)

  • fix restore OpenAI support

  • fix `analysis.ipynb`` original notebook broken with new models, evaluation for Experiment 3a (isolated) does not work

  • clean up code

Author

Please refer to the authors of the original repository

For the fork:

  • Maximilian Krupop

Back to Top