sot-phase1-tinystories-dense_half_ffn-seed123

Part of the Self-Organising Transformer research project: phase1, dataset tinystories, condition dense_half_ffn, seed 123.

Dense baseline with half the FFN width (width control).

Read this first

• Perplexity corrected on 2026-10-06: the first values came from a transformers==5.17.0 install that corrupts the reload of this model. The lm-eval scores were not re-run after that fix. At this model size they are at chance level in any case.

Full context: RESEARCH_LOG.md, the 2026-09-29 entries (runs) and the 2026-10-06 entry (corrected perplexity).

Intended use and limitations

• Intended use: reproduce and inspect the results of the Self-Organising Transformer study.
• Not for: production or downstream use. This is a small research model trained for 10,000 steps on TinyStories only. It is not instruction-tuned or safety-tuned. Its zero-shot benchmark scores are close to chance.
• It needs the project code and transformers<5.0 (see How to use).

How to use

The model class is defined in the project repository and is not registered with AutoModel, so the Hub widget and AutoModelForCausalLM do not work. Use the repository code (it needs transformers<5.0; 5.x corrupts the reload of this model):

git clone https://github.com/sbartlett97/self-organising-transformer
cd self-organising-transformer && pip install -r requirements.txt
python generate.py sbartlett97/sot-phase1-tinystories-dense_half_ffn-seed123 --prompt "Once upon a time"

from ablation_importance import resolve_checkpoint
from evaluate import load_model
from transformers import AutoTokenizer

path = resolve_checkpoint("sbartlett97/sot-phase1-tinystories-dense_half_ffn-seed123")
model = load_model(str(path), None).eval()
tok = AutoTokenizer.from_pretrained(path)
logits = model(tok("Once upon a time", return_tensors="pt").input_ids)["logits"]

Performance

evaluate.py perplexity: 100 val batches of 8 x 256 tokens.

Perplexity: 4.83

• loses to the matched dense baseline (4.61) by 4.6%

Training trajectory

• Final train loss: 1.5017 at step 10000
• Final cost proxy (gate-weighted, not compute): 1.0000
• Final training-time val loss (random val batches): 1.5613 (ppl 4.77) at step 10000

Hyperparameters