dilution-209m-16k-nope-2x-s2357
A 209M-parameter research language model from the DilutionAttention release: dilution attention in every layer, no position encoding. Trained at a 16,384-token context on 8.40B tokens (2x Chinchilla), seed 2357. It is a base model for studying the attention operator, not an assistant: it is not instruction-tuned or safety-tuned.
Results
| measure | value |
|---|---|
| validation loss (common 1M-token FineWeb-Edu slice) | 2.8541 nats/token |
| NIAH retrieval at 16,000 tokens (in context; chance 1/7) | 1.00 |
| NIAH retrieval at 24,000 tokens (1.5x the trained context) | 1.00 |
| NIAH retrieval at 32,000 tokens (2.0x the trained context) | 1.00 |
| NIAH retrieval at 48,000 tokens (2.9x the trained context) | 1.00 |
| NIAH retrieval at 64,000 tokens (3.9x the trained context) | 0.97 |
| NIAH retrieval at 96,000 tokens (5.9x the trained context) | 0.93 |
| NIAH retrieval at 128,000 tokens (7.8x the trained context) | 0.83 |
| NIAH retrieval at 192,000 tokens (11.7x the trained context) | 0.77 |
| NIAH retrieval at 256,000 tokens (15.6x the trained context) | 0.67 |
| PG-19 loss at positions 8,192-16,384 (in context) | 3.353 nats/token |
| PG-19 loss at positions 16,384-32,768 (2x) | 3.349 nats/token |
| PG-19 loss at positions 32,768-65,536 (4x) | 3.382 nats/token |
| PG-19 loss at positions 65,536-131,072 (8x) | 3.389 nats/token |
NIAH is likelihood-scored keyword retrieval from repetitive filler (see the release's reproducing.md), not long-context reasoning. Every number here, and the matched controls, are in the release's evidence ledger.
Use
pip install "dilution-attention[hub] @ git+https://github.com/Royer-Research-Labs/Dilution-Attention" tiktoken
import tiktoken, torch
from dilution.hub import from_pretrained
model = from_pretrained("Royer-Research-Labs/dilution-209m-16k-nope-2x-s2357", device="cuda") # fused Triton kernels on CUDA
enc = tiktoken.get_encoding("gpt2")
ids = torch.tensor([enc.encode("The history of the printing press")], device="cuda")
with torch.autocast("cuda", dtype=torch.bfloat16):
out = model.generate(ids, 64, temperature=1.0, top_p=0.9, vocab_limit=50257)
print(enc.decode(out[0].tolist()))
Inputs longer than 16,384 tokens need dilution.extend.ExtendedContext(model, length) around the call.
Training
- Architecture: 24 layers, width 768, 12 heads, SwiGLU, no biases, tied embeddings, GPT-2 vocabulary; position encoding: none.
- Data: FineWeb-Edu
sample-100BT(revision fc9850d), a 40B-token GPT-2-tokenized corpus rebuilt withdilution-prepare-data(reproducing.md). - Schedule: 65,536 tokens per step, LR 6e-4 cosine over a 2x-Chinchilla (8.4B-token) schedule; this is the end of that schedule.
- Run record:
runs/scale_16384/all_dilution_nope_seed2357/in the release repository (metrics, resolved config, data provenance). Source checkpoint SHA-256:f3ee60c6baba76817d80fc3850dfcdb473d1ed89c61a4b504dbdcf32f668de7a, step 128,174.
License and data
Weights: Apache-2.0. Training data: FineWeb-Edu (ODC-By 1.0), used with attribution. The model can produce incorrect, biased or offensive text; it is a research artifact.
Citation
Cite the software release: Royer, N. (2026). DilutionAttention. Zenodo. https://doi.org/10.5281/zenodo.23073037 (BibTeX in the README and CITATION.cff). Technical report: https://doi.org/10.5281/zenodo.23073151.