Sparse Model Nexus — Operational Demo
Four composition operators over three LoRA quanta · base distilgpt2 · transcript 2026-07-05T18:03:11Z
A quantum is a minimal composable LoRA adapter — the runnable form of the graph's SparseLimitModel definition. Below is a real generation transcript (greedy decode) of the ⊕ ∘ ▷ ⊗ operators composing the trained quanta. Nothing here is mocked.
⬇ Download adapter bundle (.zip)
Graph views
Trained quanta
| Quantum | Loss start | Loss end | Trainable params |
|---|
code_specialist | 3.4315 | 2.6076 | 147,456 |
biomed_specialist | 4.568 | 3.8443 | 147,456 |
shared_common | 4.5838 | 3.5852 | 147,456 |
Every quantum's loss decreased — each is genuinely specialized, not a stub.
Operator transcript (8 runs · 8.7s)
▷ route
Predicate inspects the prompt and dispatches to one specialist quantum.
prompt def reverse_linked_list(head):
chose code_specialist
reverse_linked_list()[1] = head.endof([2], 1].append(-3]) + 2, 3) return reversed_linked_list(head):
return back to top of list if not found
▷ route
Predicate inspects the prompt and dispatches to one specialist quantum.
prompt The function of insulin in glucose metabolism is
chose code_specialist
to increase the amount of blood sugar. The effect on this happens when a person has an increased number of calories, and that's because it increases their body fat by increasing its own weight (which means they're more likely than not).
▷ route
Predicate inspects the prompt and dispatches to one specialist quantum.
prompt Here is a short note about the weather:
chose shared_common
The sun will be shining over this afternoon. The sky may well rise in some parts of the country, but it does not appear to have been bright enough for most people on Sunday morning (except at night).
⊕ mixture
Several quanta blended in weight-space at chosen widths (MoE-style). A 0 weight is the null-expert identity.
prompt Explain the following clearly:
weights code_specialist=1, shared_common=0
The first thing you need to do is make sure that your computer has a USB port. If it's not, then try again and connect with this one of these ports on both sides of the screen (or if there are two).
⊕ mixture
Several quanta blended in weight-space at chosen widths (MoE-style). A 0 weight is the null-expert identity.
prompt Explain the following clearly:
weights code_specialist=0.5, shared_common=0.5
The first thing you notice is that there are two main points. The second point, which I will describe in a later section of this article (which may be more detailed), and the third one has an important distinction between what we see
⊕ mixture
Several quanta blended in weight-space at chosen widths (MoE-style). A 0 weight is the null-expert identity.
prompt Explain the following clearly:
weights code_specialist=0.7, biomed_specialist=0.3
The first thing you need to do is make sure that your device has a USB port. If it's not, then try again and connect with another PC or Mac before connecting via Ethernet (or other means). This will be done by
∘ sequential
Quanta chained — each one's output is threaded into the next (non-commutative).
chain shared_common ∘ code_specialist
shared_commonThe following code is provided for the purposes of this section:
code_specialistThe function that returns an array with its first argument, and then takes two arguments to return it (the last one being passed). If we want to use both arrays as input or output in our
⊗ merge
Task-arithmetic merge of several quanta into one new deployable quantum.
merge shared_common ⊗ code_specialist ⊗ biomed_specialist → omni_quantum
The main problem is that the first step in explaining why a given question has no answer. The second issue of this series will be about how to explain it and what we can do with them (see below). In particular I'll discuss
Reproduce locally
pip install torch transformers peft
# unzip sparse-model-nexus-adapters.zip, then:
python demo.py # regenerates this transcript
python -c "from sparse_nexus import Nexus; print(Nexus().route('def quicksort(xs):')['output'])"