data-similarity-is-not-enough

Code and data for "Data Similarity is Not Enough to Explain Language Model Performance" (EMNLP 2023).

Requirements

The file requirements.txt contains the Python packages required to run this code. This code is written for use with a GPU.

Raw results

This repository contains the raw similarity and performance results used to construct the figures and tables in the paper:

• raw-results/bigbench-raw-results-multiple-choice.csv: overall BIG-bench Lite performances and similarities used to create the figures and tables
• raw-results/cosine-similarities: This directory contains the cosine similarities between Sentence-T5 embeddings of sampled downstream documents and all of the Pile and C4.
• raw-results/sampled-pretraining-idxs: This directory contains the indices of sampled pretraining documents from the Pile and C4. These correspond to indices in the datasets when loaded using Hugging Face datasets. They were used for calculating KL-divergences and MAUVE scores against samples of the pretraining datasets.

The raw-results directory also contains the performances and similarities for Stack Exchange and XNLI datasets.

Reproducing figures and tables from the paper with provided results

To reproduce the figures and tables from the paper, run the following commands:

• python tables-1-and-3.py: generates the tables with correlations between similarity and performance for BIG-bench Lite multiple choice tasks
• python table-2.py: generates the table with correlations between similarity measures when comparing BIG-bench Lite with C4 and the Pile
• python figure-1.py --reproduction: generates the graphs of similarity and performance for Stack Exchange and BIG-bench Lite multiple choice tasks.
• python figure-2.py --reproduction: generates the graphs of similarity and performance for Stack Exchange and XNLI.
• python figure-bigbench-violinplots.py --model Pythia-6.9B: generates the BIG-bench Lite violinplots in figure 3
• python figure-4.py: generates the GLUE violinplots in figure 4
• python figure-5.py --reproduction: generates small multiples for all XNLI languages
• python figure-6.py: generates the BIG-bench Lite barplots binned by similarity quartile
• python figure-bigbench-violinplots.py --model T5: generates the BIG-bench Lite violinplots in figure 7
• python figure-8.py: generates the Stack Exchange AM/PM graph of similarity and performance

---

Datasets

Stack Exchange classification

Each file contains 1000 posts from Stack Exchange: 500 posts from each of the bicycles and cstheory Stack Exchange forums. All posts were originally in English. The percentage in the filename refers to how much of each post has been translated by Google Translate. For example, the file bicycles-cstheory_50-percent-finnish.json contains posts whose first half is in Finnish. The original dataset is in bicycles-cstheory_100-percent-english.json.

These datasets were adapted from Comparing Text Representations: A Theory-Driven Approach (EMNLP 2021).

XNLI

Each file contains 2,500 natural language inference questions. All documents were originally in English. Again, the percentage in the filename refers to how much of each document has been translated by Google Translate. For example, the file xnli_50-percent-spanish contains documents whose first half is in Spanish. The original XNLI datasets are those that have 100-percent in the filename.

These datasets were adapted from XNLI: Evaluating cross-lingual sentence representations (EMNLP 2018).

Prompt templates

Stack Exchange forum classification: Each example contains the post text and the forum it was posted to (either 'bicycles' or 'cstheory'). We format each example as: {text}\nThis post is about {forum}. The prompt is the same across languages.

Stack Exchange AM/PM classification: Each example contains the post text and whether it was posted in the 'morning' or 'afternoon'. We format each example as: {text}\nThis was posted in the {label}. The prompt is the same across languages.

XNLI: Since XNLI consists of translations of MNLI, we use an MNLI prompt for all datasets. Examples include a premise, a hypothesis, and a label. Labels are mapped, in order, from 'entailment', 'neutral', 'contradiction' to 'Yes', 'Maybe', 'No'. An example then becomes: {premise} {label}, {hypothesis}. The prompt is the same across languages.

BIG-bench Lite: BIG-bench Lite examples come pre-formatted when loaded through the Hugging Face datasets library. We use the existing formatting.

GLUE: For T5 experiments, we use the prompting style described in the original T5 paper (Raffel et al., JMLR 2020). For example, an MNLI example with a premise and a hypothesis becomes: 'mnli hypothesis: {hypothesis} premise: {premise} {label}' where the label can be 'entailment', 'contradiction', or 'neutral'.

---

Code

This code calculates zero-shot accuracy on BIG-bench Lite, GLUE, Stack Exchange, and XNLI. It then calculates each downstream dataset's similarity to two pretraining datasets (the Pile and C4).

1. Download pretraining datasets. Sample them and create token distributions and ST5 embeddings.

sample_pretraining_data.py: Downloads C4 and the Pile, samples 100,000 docs at a time, and creates ST5 embeddings. Arguments:

• dataset_name: can be either the_pile or c4
• sample_num: an integer identifier for this sample--we use integers from 1 to 8
• num_samples: the number of documents to sample

n-gram-distributions-pretraining.py: Generate token distributions for samples from C4 and the Pile. Arguments:

• dataset_name: can be either the_pile or c4
• sample_num: an integer identifier for this sample--we use integers from 1 to 8

2. Evaluate downstream datasets

eval-stackexchange.py: Run Pythia-6.9B zero-shot eval on Stack Exchange classification tasks. Arguments:

• label_type: Can be either forum or ampm.
• model: Name of a transformers model. We report results with Pythia-6.9B.

eval-xnli.py: Run Pythia-6.9B zero-shot eval on XNLI classification tasks. Arguments:

• model: Name of a transformers model. We report results with Pythia-6.9B.

eval-bigbench.py: Run Pythia-6.9B zero-shot and few-shot eval on BIG-bench Lite tasks. Arguments:

• model: Name of a transformers model. We report results with Pythia-6.9B.

eval-bigbench-t5.py: Run T5-3B and T5 v1.1 XL zero-shot and few-shot eval on BIG-bench Lite tasks. Arguments:

• model: Name of a transformers model. Options are t5 for finetuned T5-3B and t5-v1_1-xl for T5 v1.1 XL.

eval-glue.py: Run zero-shot eval on GLUE tasks.

• model: Name of a transformers model.

3. Get token distributions and embeddings for downstream tasks.

Run the following files to create token distributions:

• n-gram-distributions-bigbench.py
• n-gram-distributions-stackexchange.py
• n-gram-distributions-xnli.py

Run the following file to construct embeddings:

• embeddings-bigbench.py
• embeddings-glue.py.

4. Process downstream results and calculate similarities.

Run the following files:

• process-stackexchange.py
• process-xnli.py
• process-bigbench.py
• process-glue.py

We are unable to include our code for calculating cosine similarities between entire pretraining datasets and examples from downstream datasets, but the scripts in this section calculate such similarities against a sample of the pretraining dataset. We include our raw results against the entire pretraining dataset in the cosine-similarities directory.

5. Make figures

Run the following files to generate the tables and figures in the paper:

• tables-1-and-3.py: Generates tables 1 and 3.
• table-2.py: Generates tables 2.
• figure-1.py: Generates figure 1.
• figure-2.py: Generates figure 2.
• figure-bigbench-violinplots.py: Generates figures 3 and 7. Arguments are model (either Pythia-6.9B or T5) and similarity_type (either max or mean).
• figure-4.py: Generates figure 4.
• figure-5.py: Generates figure 5.
• figure-6.py: Generates figure 6. Arguments are model (either Pythia-6.9B or T5).
• process-stackexchange-ampm.py: Generates figure 8.

---

Raw performance results

This table contains the raw performance results for BIG-bench Lite multiple choice tasks that are used to construct tables and figures but that are not explicitly included in the main paper.

Dataset	Pythia-6.9B (0 shot)	T5 v1.1 XL (0 shot)	T5 v1.1 XL (2 shot)	Flan-T5 XL (0 shot)	Flan-T5 XL (2 shot)
bbq_lite_json	-22.02	4.80	8.75	19.90	37.76
code_line_description	-2.00	-10.86	-10.86	17.96	24.61
conceptual_combinations	0.00	0.32	5.50	52.10	41.75
emoji_movie	6.25	-1.25	-5.00	3.75	3.75
formal_fallacies_syllogisms_negation	-0.07	0.00	0.00	1.18	2.89
hindu_knowledge	1.96	-0.51	1.02	14.72	16.24
known_unknowns	4.35	0.00	4.35	-8.70	-17.39
language_identification	5.49	0.25	-0.08	12.34	4.73
logic_grid_puzzle	18.88	-2.60	-2.60	8.43	0.01
logical_deduction	0.00	-1.04	0.16	15.91	11.09
novel_concepts	-8.53	-21.09	-17.19	-9.38	-1.56
operators	0.00	0.95	0.00	5.71	6.19
play_dialog_same_or_different	21.94	-26.16	-26.16	18.75	15.63
strange_stories	-3.02	-9.56	-5.79	46.16	34.83
strategyqa	3.89	-6.42	-6.51	7.56	28.35
symbol_interpretation	-1.01	-0.51	-2.40	1.01	1.01
vitaminc_fact_verification	24.57	5.34	5.34	56.71	41.31