sot-phase2-wikitext103-heuristic-seed42-tc768d013-v8-tailfreeze

Part of the Self-Organising Transformer research project: phase2, dataset wikitext103, condition heuristic, seed 42.

Phase 2: training starts with 33% of units alive; a controller gives birth to dormant units and kills weak ones every 250 steps.

Read this first

• Phase 2 attempt 8.
• Every number on this card is a supernet number. This run predates the zero-effect fix (2026-10-15): dormant units (gate about 0.04) and dead units (gate about 0.001) were still in the forward pass and still trained. The checkpoint loads with every scale at 1, so it runs as trained.
• Births and deaths frozen below 5% of peak LR: val ppl falls at every eval to 26.98.
• The present structure alone writes degenerate text (samples).

Full context: RESEARCH_LOG.md, the 2026-10-14 attempt-8 and compute-accounting entries.

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 WikiText-103 only. It is not instruction-tuned or safety-tuned.
• 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-phase2-wikitext103-heuristic-seed42-tc768d013-v8-tailfreeze --prompt "The history of the city begins in"

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

path = resolve_checkpoint("sbartlett97/sot-phase2-wikitext103-heuristic-seed42-tc768d013-v8-tailfreeze")
model = load_model(str(path), None).eval()
tok = AutoTokenizer.from_pretrained(path)
logits = model(tok("The history of the city begins in", return_tensors="pt").input_ids)["logits"]

Performance

No evaluate.py run. The final training-time val ppl of the supernet is in "Training trajectory" below. Phase 1 dense 768 on the same method: 24.30.

Val ppl on 48 val batches of 8 x 256 (structure_eval.py):

Final present structure

9 of 12 layers present: 0.466 of dense FLOPs/token, 0.472 of dense parameters (compute.py). The checkpoint itself still holds every unit.

Structural plasticity events

• Total births: 92
• Total deaths: 23
• Total structural events: 115

Training trajectory

• Final train loss: 3.2271 at step 10000
• Final cost proxy (gate-weighted, not compute): 0.0530
• Final training-time val loss (random val batches): 3.2952 (ppl 26.98) at step 10000

Hyperparameters