Custodial Weights — model snapshot
Read this first. The point of Custodial Weights is not this frozen file. The point is the mechanism: a model whose weights are derived from a knowledge graph and updated in memory, continuously, as it learns. This is a snapshot of that dynamic state, at rest. Downloading it is like photographing a river.
The idea, in one paragraph
Most language models train once, freeze, and never change after release. This project removes that assumption. The weights here are derived from a knowledge graph and updated as the model learns; a Gate decides what counts as true and rejects contradictions before they reach the weights; the model answers from what it believes at that moment, because talking changes the weights.
This model owns its own weights. That is a step toward a different kind of thing than a frozen autocomplete — plainly stated, honestly hedged. This is not a claim of consciousness. It is a mechanism for self-custody.
The checkpoint
st4c3.final.pt — the headline run:
- 4.17M parameters (RMSNorm, RoPE, GQA, SwiGLU;
d_model=256, n_layer=4) - Trained on the expanded kinship dataset with heavy contrast curriculum
- Relational verification: top-1 0.830, top-3 1.000 on tested facts (chance top-1 ≈ 0.111)
How to use it (honest)
This is not a drop-in from_pretrained model. To reproduce the verified
relational behavior you need the full pipeline (Gate + tokenizer + family
index), because verification is a relative ranking within a family, not an
absolute score. That pipeline is in the
code repository, along with
src/, scripts/eval.py, and the evaluation data:
git clone https://github.com/n00beroUno/Custodial-Weights
cd Custodial-Weights
pip install -r requirements.txt
python scripts/eval.py # print the published numbers
To re-evaluate this exact checkpoint, download st4c3.final.pt from this
model repo into the repo root and run:
python scripts/eval.py ./st4c3.final.pt
(The full-scale evaluation needs the converted dataset that is gitignored
from the code repo; see docs/dataset.md for how to regenerate it.)
You'll also need the training dataset.
What the numbers mean, honestly
| checkpoint | top-1 | top-3 |
|---|---|---|
| st3b (no contrast) | 0.335 | 0.893 |
| st4c4 (light contrast) | 0.502 | 0.998 |
| st4c3 (heavy contrast) | 0.830 | 1.000 |
| st4c5 (heavy contrast, 6.44M) | 0.797 | 1.000 |
Relational verify improves monotonically with contrast dose and capacity is not the lever — doubling the model (4.17M → 6.44M) did not help. The data and the curriculum are what matter.
Try it live
Run the whole mechanism in your browser, no download: Custodial Weights demo Space.