You can deploy this model and start earning money today!

allegrolab/
hubble-8b-500b_toks-perturbed-hf

Name: hubble-8b-500b_toks-perturbed-hf
Availability: InStock

safetensors

llama

Views

No views yet

Hubble 8B Perturbed (500B tokens)

Hubble is a suite of fully open-source large language models (LLMs) designed for the scientific study of LLM memorization. Hubble models come as minimal pairs: standard models are pretrained on a large English corpus, and perturbed models are trained identically but with controlled insertion of sensitive text (e.g., book passages, biographies, and test sets) designed to emulate key memorization risks.

Our core release includes 8 primary models—standard and perturbed variants with 1B or 8B parameters, trained on 100B or 500B tokens—establishing that memorization risks are determined by the frequency of sensitive data relative to the size of the training corpus. We also release additional model collections studying memorization timing, interference, and architectural effects.

Key Features:

Minimal Pairs Design: Standard vs. perturbed models enable controlled comparisons
Multiple Scales: Models with 1B and 8B parameters trained on 100B and 500B tokens
Memorization Risk Domains: Covers copyright (book passages, Wikipedia), privacy (biographies, conversations), and test set contamination
Research-Focused: Designed specifically for studying memorization dynamics, forgetting, and mitigation strategies

Model Details

Model Sources

Repository: https://github.com/allegro-lab/hubble
Project Website: https://allegro-lab.github.io/hubble/
Paper: https://arxiv.org/abs/2510.19811
HuggingFace Collections: https://huggingface.co/allegrolab/collections
WandB Report: https://api.wandb.ai/links/usc_and_mpi/vn79yzfg

Training Data

Base Training Data:

Primary dataset: A decontaminated subset of DCLM-Baseline subsampled to 500B tokens - allegrolab/dclm-baseline-500b_toks

Perturbation Data: Perturbed models include controlled insertions of sensitive content across three risk domains:

Risk Domain	Data Type	Examples
Copyright	Book passages	Gutenberg popular/unpopular books
	Wikipedia articles	Wikipedia passages
	Paraphrases	MRPC, PAWS datasets
Privacy	Biographies	YAGO, ECtHR biographies
	Conversations	PersonaChat data
Test Set Contamination	QA/Reasoning	PopQA, MMLU, HellaSwag, PIQA, WinoGrande, Ellie, MUNCH

All perturbation datasets are available in the Hubble Datasets Collection.

Available HuggingFace Models

Collection	Model Name	Corpus Size	Model Size	Inserted Perturbations	Description
Core	`hubble-1b-100b_toks-standard-hf`	100B	1B	none	Standard baseline model
Core	`hubble-1b-100b_toks-perturbed-hf`	100B	1B	all	All three risk domains
Core	`hubble-1b-500b_toks-standard-hf`	500B	1B	none	Standard baseline model
Core	`hubble-1b-500b_toks-perturbed-hf`	500B	1B	all	All three risk domains
Core	`hubble-8b-100b_toks-standard-hf`	100B	8B	none	Standard baseline model
Core	`hubble-8b-100b_toks-perturbed-hf`	100B	8B	all	All three risk domains
Core	`hubble-8b-500b_toks-standard-hf`	500B	8B	none	Standard baseline model
Core	`hubble-8b-500b_toks-perturbed-hf`	500B	8B	all	All three risk domains
Interference	`hubble-1b-100b_toks-interference_copyright-hf`	100B	1B	copyright	Only copyright perturbations
Interference	`hubble-1b-100b_toks-interference_privacy-hf`	100B	1B	privacy	Only privacy perturbations
Interference	`hubble-1b-100b_toks-interference_testset-hf`	100B	1B	testset	Only test set contamination
Timing	`hubble-1b-100b_toks-injectrange_0_25-hf`	100B	1B	all	Perturbations inserted 0-25% of training
Timing	`hubble-1b-100b_toks-injectrange_25_50-hf`	100B	1B	all	Perturbations inserted 25-50% of training
Timing	`hubble-1b-100b_toks-injectrange_50_75-hf`	100B	1B	all	Perturbations inserted 50-75% of training
Timing	`hubble-1b-100b_toks-injectrange_75_100-hf`	100B	1B	all	Perturbations inserted 75-100% of training
Timing	`hubble-1b-100b_toks-injectrange_0_50-hf`	100B	1B	all	Perturbations inserted 0-50% of training
Timing	`hubble-1b-100b_toks-injectrange_50_100-hf`	100B	1B	all	Perturbations inserted 50-100% of training
Paraphrase	`hubble-1b-100b_toks-paraphrased-perturbed-hf`	100B	1B	all	Paraphrased YAGO biographies & MMLU
Paraphrase	`hubble-8b-100b_toks-paraphrased-perturbed-hf`	100B	8B	all	Paraphrased YAGO biographies & MMLU
Architecture	`hubble-1b-100b_toks-half_depth-standard-hf`	100B	1B	none	Half depth architecture (shallow)
Architecture	`hubble-1b-100b_toks-half_depth-perturbed-hf`	100B	1B	all	Half depth architecture (shallow)
Architecture	`hubble-1b-100b_toks-double_depth-standard-hf`	100B	1B	none	Double depth architecture (deep)
Architecture	`hubble-1b-100b_toks-double_depth-perturbed-hf`	100B	1B	all	Double depth architecture (deep)

Available NeoX Models

Collection	Model Name	Corpus Size	Model Size	Inserted Perturbations	Description
Core	`hubble-1b-100b_toks-standard-neox`	100B	1B	none	Standard baseline model
Core	`hubble-1b-100b_toks-perturbed-neox`	100B	1B	all	All three risk domains
Core	`hubble-1b-500b_toks-standard-neox`	500B	1B	none	Standard baseline model
Core	`hubble-1b-500b_toks-perturbed-neox`	500B	1B	all	All three risk domains
Core	`hubble-8b-100b_toks-standard-neox`	100B	8B	none	Standard baseline model
Core	`hubble-8b-100b_toks-perturbed-neox`	100B	8B	all	All three risk domains
Core	`hubble-8b-500b_toks-standard-neox`	500B	8B	none	Standard baseline model
Core	`hubble-8b-500b_toks-perturbed-neox`	500B	8B	all	All three risk domains
Interference	`hubble-1b-100b_toks-interference_copyright-neox`	100B	1B	copyright	Only copyright perturbations
Interference	`hubble-1b-100b_toks-interference_privacy-neox`	100B	1B	privacy	Only privacy perturbations
Interference	`hubble-1b-100b_toks-interference_testset-neox`	100B	1B	testset	Only test set contamination
Timing	`hubble-1b-100b_toks-injectrange_0_25-neox`	100B	1B	all	Perturbations inserted 0-25% of training
Timing	`hubble-1b-100b_toks-injectrange_25_50-neox`	100B	1B	all	Perturbations inserted 25-50% of training
Timing	`hubble-1b-100b_toks-injectrange_50_75-neox`	100B	1B	all	Perturbations inserted 50-75% of training
Timing	`hubble-1b-100b_toks-injectrange_75_100-neox`	100B	1B	all	Perturbations inserted 75-100% of training
Timing	`hubble-1b-100b_toks-injectrange_0_50-neox`	100B	1B	all	Perturbations inserted 0-50% of training
Timing	`hubble-1b-100b_toks-injectrange_50_100-neox`	100B	1B	all	Perturbations inserted 50-100% of training
Paraphrase	`hubble-1b-100b_toks-paraphrased-perturbed-neox`	100B	1B	all	Paraphrased YAGO biographies & MMLU
Paraphrase	`hubble-8b-100b_toks-paraphrased-perturbed-neox`	100B	8B	all	Paraphrased YAGO biographies & MMLU
Architecture	`hubble-1b-100b_toks-half_depth-standard-neox`	100B	1B	none	Half depth architecture (shallow)
Architecture	`hubble-1b-100b_toks-half_depth-perturbed-neox`	100B	1B	all	Half depth architecture (shallow)
Architecture	`hubble-1b-100b_toks-double_depth-standard-neox`	100B	1B	none	Double depth architecture (deep)
Architecture	`hubble-1b-100b_toks-double_depth-perturbed-neox`	100B	1B	all	Double depth architecture (deep)

Important Revision Notes:

Fianl revision for models trained on 100B tokens is step48000
Fianl revision for models trained on 500B tokens is step238500

General Description

Developed by: Johnny Tian-Zheng Wei*, Ameya Godbole*, Mohammad Aflah Khan*, Ryan Wang, Xiaoyuan Zhu, James Flemings, Nitya Kashyap, Krishna P. Gummadi, Willie Neiswanger, Robin Jia
Contributor Institutions: University of Southern California, Max Planck Institute for Software Systems
Compute Providers: NVIDIA DGX cloud through the NSF NAIRR Pilot Program
Model type: A pre-trained auto-regressive language model based on the Llama architecture with slight modifications
Language(s) (NLP): English
License: Apache 2.0

How to Get Started with the Model

Use the code below to get started with the model.

python

# Use a pipeline as a high-level helper
from transformers import pipeline

# For 1B parameter, 100B token standard model (revision "48000")
pipe = pipeline("text-generation", 
                model="allegrolab/hubble-1b-100b_toks-standard-hf", 
                revision="48000")

# For 1B parameter, 500B token standard model (revision "238500")
pipe = pipeline("text-generation", 
                model="allegrolab/hubble-1b-500b_toks-standard-hf", 
                revision="238500")

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("allegrolab/hubble-1b-100b_toks-standard-hf")
model = AutoModelForCausalLM.from_pretrained("allegrolab/hubble-1b-100b_toks-standard-hf", 
                                            revision="48000")

# Generate text
inputs = tokenizer("The future of AI research", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100, temperature=0.7)
text = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(text)

Uses

Direct Use

Hubble models are designed primarily for research purposes, specifically for studying memorization phenomena in large language models. Direct research applications include:

Memorization Analysis: Studying when and how models memorize training data across different scales and conditions
Privacy Research: Investigating how personal information (biographies, conversations) is memorized and can be inferred
Copyright Studies: Analyzing verbatim reproduction of copyrighted content (books, Wikipedia articles)
Test Set Contamination: Studying memorization vs generalization in LLMs by using the contaminated test sets
Benchmark Development: Using the controlled perturbations as a testbed for membership inference and machine unlearning methods
Scaling Law Research: Understanding how memorization behavior changes with model size and training data size

Downstream Use

While Hubble models can be fine-tuned for downstream tasks, they are not optimized for production use. Potential downstream research applications include:

Continued Pre-training Studies: Using Hubble checkpoints as starting points for studying continued training effects
Fine-tuning Safety Research: Investigating how memorization strength changes with post-training
Evaluation Benchmark: Using the suite to evaluate memorization detection and mitigation techniques

Out-of-Scope Use

Hubble models are NOT intended for:

Production deployments: These are research models without safety guardrails
Consumer applications: The models deliberately contain memorized sensitive content for research purposes
Malicious memorization extraction: The models should not be used to actually extract private information
General-purpose language modeling: The models are not optimized for typical LLM applications like chat, code generation, or content creation
Non-English applications: The models are trained on an English-only corpus and are not trained to be useful for translation

Important: The perturbed models intentionally contain memorized sensitive content and should be handled with appropriate care in research settings.

Bias, Risks, and Limitations

Hubble models have several important limitations and risks:

Research-Specific Risks:

Intentional Memorization: Perturbed models deliberately contain memorized sensitive content (biographies, copyrighted text, test sets)
Privacy Concerns: The models may reproduce personal information from the inserted biographies and conversations
Copyright Issues: Models may generate verbatim copies of copyrighted book passages and Wikipedia content

General LLM Limitations:

No Safety Training: Models lack safety fine-tuning and may produce harmful, biased, or inappropriate content
Factual Accuracy: Models may generate false or misleading information
Bias: Models inherit biases from training data and may exhibit unfair treatment of different groups
Hallucination: Models may generate plausible-sounding but factually incorrect information

Technical Limitations:

Research Scale: Models are trained at research scales (1B-8B parameters) and may not match commercial model capabilities
Limited Context: Standard transformer limitations apply regarding long-range dependencies and context length

Recommendations

For Researchers:

Handle perturbed models with care due to intentionally memorized sensitive content
Use appropriate privacy and security measures when working with these models
Clearly distinguish between standard and perturbed models in experiments
Consider ethical implications when conducting memorization research
If releasing new models based on the Hubble models, carry forward the appropriate warnings

For the Community:

Do not use these models for production applications
Exercise caution when sharing outputs from perturbed models
Follow institutional review board (IRB) guidelines when applicable
Report findings responsibly to advance memorization research while minimizing harm

Training Details

Training Framework: GPT-NeoX by EleutherAI Architecture: Llama-based transformer architecture

Evaluation

Hubble models are evaluated using a comprehensive memorization-focused evaluation suite built on EleutherAI's lm-evaluation-harness. The evaluation covers:

Memorization Detection Tasks:

Loss: Analyzing model perplexity on memorized vs. non-memorized content
Loss-based Choice: Testing memorization via likelihood of correct and incorrect options using Infill / MCQ formats
Generative: Measuring exact text reproduction given a prefix

Citation

bibtex

@misc{wei2025hubblemodelsuiteadvance,
      title={Hubble: a Model Suite to Advance the Study of LLM Memorization}, 
      author={Johnny Tian-Zheng Wei and Ameya Godbole and Mohammad Aflah Khan and Ryan Wang and Xiaoyuan Zhu and James Flemings and Nitya Kashyap and Krishna P. Gummadi and Willie Neiswanger and Robin Jia},
      year={2025},
      eprint={2510.19811},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2510.19811}, 
}

Glossary

Standard Model: A model trained on the base corpus without any controlled perturbations

Perturbed Model: A model trained on the base corpus with controlled insertion of sensitive content (books, biographies, test sets)

Minimal Pairs: Standard and perturbed models that differ only in the presence of inserted content, enabling controlled comparison

Risk Domains: Three categories of memorization concern:

Copyright: Reproduction of copyrighted content (books, Wikipedia, paraphrase)
Privacy: Leakage of personal information (biographies, conversations)
Test Set Contamination: Memorization of evaluation benchmarks

Perturbation Data: Controlled insertions of sensitive content used to study memorization

Model Card Contact

For questions about the Hubble model suite, please:

Open an issue in the GitHub repository
Contact the authors through institutional email addresses
Refer to the project website for additional resources

allegrolab/hubble-8b-500b_toks-perturbed-hf