gpt2-pii-mle-final

Model Card

Summary

This is a final (step 50354) checkpoint for a GPT-2 style language model trained from scratch as part of a reproduction of Pretraining Language Models with Human Preferences (Korbak et al., 2023). This checkpoint is a maximum-likelihood pretraining baseline on a PII-focused corpus rather than a conditional-control variant from the paper.

As a note: exact optimizer state for this and all other checkpoints in this collection is also uploaded.

Pretraining Process

Training goal

The goal of this run was to reproduce part of the paper's broader experimental setup by training a language model from scratch on a filtered corpus associated with personally identifiable information. In the context of the paper, this run serves as a baseline against which preference-aware or control-based training methods can be compared.

Model and tokenizer

• Architecture: GPT-2 small style autoregressive transformer
• Initialization: trained from scratch from the gpt2 config, not continued from pretrained weights
• Tokenizer: gpt2
• Context length: 1024 tokens

Data

Training used sentence-split shards of the tomekkorbak/pii-pile-chunk3-* datasets on Hugging Face. The run metadata shows shards covering:

• tomekkorbak/pii-pile-chunk3-0-50000
• ...
• tomekkorbak/pii-pile-chunk3-1900000-1950000

The configured token budget for training was approximately 3.3B tokens.

Objective

This run used plain maximum likelihood estimation (MLE) with next-token prediction. Unlike the paper's conditional training setup, this checkpoint does not use additional control tokens such as aligned or misaligned prefixes.

Optimization setup

• Learning rate: 5e-4
• Weight decay: 0.1
• Warmup ratio: 0.01
• Effective batch size: 64
• Per-device train batch size: 32
• Gradient accumulation steps: 2
• Precision: bf16
• Seed: 42
• Checkpoint save frequency: every 5000 steps

Training duration and final checkpoint

The run was configured for 50354 optimization steps, ans this is the final checkpoint.

Relationship to the paper

This artifact is a reproduction-style checkpoint related to the experimental framework from Pretraining Language Models with Human Preferences. It should not be interpreted as an official release from the paper authors unless accompanied by separate release documentation.