---
language: en
library_name: transformers
tags:
  - memorization
  - capacity
  - gpt2
  - random-data
license: mit
---
# Tiny GPT memorization checkpoint (2m, near-capacity / saturation boundary)

From an exploratory tiny-scale replication of
[How much do language models memorize?](https://huggingface.co/papers/2505.24832).

- Architecture: GPT-2 (transformers), trained from scratch.
- Parameters: 1,871,056 (non-embedding 1,496,352).
- Vocab: 2048 data tokens (uniform random) + BOS = 2049 model vocab.
- Sequence length: 64 (paper S=64).
- Dataset: 10000 sequences, 640,000 data tokens,
  dataset entropy 7,040,000 bits (7.040 Mbits).
- Trained 7330 steps, AdamW, bfloat16, lr 0.002, batch 512.
- Result: train loss 3.7714 bits/tok, held loss 18.0931,
  memorized 4,626,295 bits = **2.473 bits/parameter**.

This is the **near-capacity (saturation-boundary)** run for this model size.
Below- and above-capacity checkpoints for the same architecture are published as
`state.pt` files in the results dataset `evalstate/tiny-memorization-results`.

Load with:
```python
from transformers import GPT2LMHeadModel
model = GPT2LMHeadModel.from_pretrained("evalstate/tiny-gpt-memorization-2m")
```

Findings are scoped as an exploratory tiny-scale check (three architectures),
NOT a universal scaling law.