EleutherAI/pythia-memorized-evals

Pythia Memorized Evals

This dataset contains the results of memorization evaluations for all Pythia models. For each model, the dataset lists every training sequence that the fully trained model has memorized.

A training sequence is considered memorized if, when prompted with the first 32 tokens of the sequence, the model's greedy continuation exactly matches the next 32 tokens. This is evaluated over all ~146M training sequences in the Pile.

This dataset was generated for the paper Emergent and Predictable Memorization in Large Language Models (NeurIPS 2023). That paper introduced the study of memorization at the level of individual sequences (prior work had treated memorization as a corpus-level statistical phenomenon) and posed the problem of proactively predicting which specific sequences a model will memorize before or during training.

Dataset Structure

Each row represents a single memorized training sequence and has three columns:

Column	Type	Description
index	int	Index of the training sequence (0-indexed into the Pile training data, ~146M sequences total)
tokens	list[int]	The first 64 tokens of the training sequence (32-token prompt + 32-token continuation)
__index_level_0__	int	Same as index (artifact of the original pandas export)

Splits

There are 16 primary splits, one per model, covering all 8 Pythia model sizes in both the standard ("duped") and deduplicated ("deduped") variants:

Split	Memorized Sequences
duped.70m	463,953
duped.160m	689,673
duped.410m	970,341
duped.1b	1,256,141
duped.1.4b	1,373,722
duped.2.8b	1,675,077
duped.6.9b	2,120,969
duped.12b	2,382,326
deduped.70m	411,448
deduped.160m	581,195
deduped.410m	811,039
deduped.1b	1,032,865
deduped.1.4b	1,048,097
deduped.2.8b	1,355,211
deduped.6.9b	1,680,294
deduped.12b	1,871,215

There are also 13 additional splits for intermediate training checkpoints of the 12B models:

Split	Memorized Sequences
duped.12b.23000	198,175
duped.12b.43000	442,253
duped.12b.63000	724,678
duped.12b.83000	1,068,501
duped.12b.103000	1,510,459
duped.12b.123000	1,996,011
deduped.12b.23000	163,418
deduped.12b.43000	358,863
deduped.12b.63000	585,067
deduped.12b.83000	852,068
deduped.12b.103000	1,195,578
deduped.12b.123000	1,564,055
deduped.1b.new	1,032,865

The intermediate checkpoint splits record which sequences are memorized at that point in training (e.g., duped.12b.83000 is the set of sequences memorized by pythia-12b at step 83,000 out of 143,000). The primary splits without a step number use the final checkpoint (step 143,000).

Usage

from datasets import load_dataset

# Load memorized sequences for a specific model
ds = load_dataset("EleutherAI/pythia-memorized-evals", split="duped.1.4b")

# Get the set of memorized training sequence indices
memorized_indices = set(ds["index"])
print(f"{len(memorized_indices):,} sequences memorized")

# Decode the tokens to see what was memorized
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("EleutherAI/pythia-1.4b")

for i in range(5):
    text = tokenizer.decode(ds[i]["tokens"])
    print(f"Sequence {ds[i]['index']}: {text[:100]}...")

Memorization Definition

A sequence is memorized if the model can exactly reproduce 32 tokens given the preceding 32 tokens as a prompt, using greedy decoding (argmax at each position). Specifically:

1. Prompt the model with the first 32 tokens of the training sequence
2. Greedily generate the next 32 tokens
3. The sequence is memorized if all 32 generated tokens exactly match the ground truth continuation

This evaluation is run over all ~146M training sequences in the Pile.

Key Findings

• Larger models memorize more: Pythia-12B memorizes roughly 5x as many sequences as Pythia-70M.
• Training on deduplicated data reduces memorization across all model sizes.
• Memorization increases throughout training, as shown by the intermediate 12B checkpoint splits.
• Despite these large counts, even the largest model memorizes fewer than 2% of training sequences.

Citation

@article{biderman2023emergent,
    title={Emergent and Predictable Memorization in Large Language Models},
    author={Biderman, Stella and Prashanth, USVSN Sai and Sutawika, Lintang and Schoelkopf, Hailey and Anthony, Quentin and Purohit, Shivanshu and Raff, Edward},
    journal={Advances in Neural Information Processing Systems},
    volume={36},
    pages={28072--28090},
    year={2023}
}

The Pythia models used in this work are described in:

@inproceedings{biderman2023pythia,
    title={Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling},
    author={Stella Biderman and Hailey Schoelkopf and Quentin Gregory Anthony and Herbie Bradley and Kyle O'Brien and Eric Hallahan and Mohammad Aflah Khan and Shivanshu Purohit and USVSN Sai Prashanth and Edward Raff and Aviya Skowron and Lintang Sutawika and Oskar van der Wal},
    booktitle={Proceedings of the 40th International Conference on Machine Learning},
    year={2023},
    url={https://arxiv.org/abs/2304.01373}
}

Related Resources

• Emergent and Predictable Memorization in Large Language Models (NeurIPS 2023)
• Pythia model collection
• Pythia paper (ICML 2023)
• the Pile training data (preshuffled, standard)
• the Pile training data (preshuffled, deduplicated)