Add oracle definition — explain the naming and oracle-soul architecture philosophy
Browse files
README.md
CHANGED
|
@@ -30,6 +30,16 @@ STEM-Oracle shares the oracle-soul architecture with [Opus-Candid-27B-V3.5](http
|
|
| 30 |
|
| 31 |
---
|
| 32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
## What Makes This Different from Math-Distilled Models
|
| 34 |
|
| 35 |
Standard STEM fine-tunes (WizardMath, MetaMath, etc.) train on solution-answer pairs. The model learns to pattern-match problem structures to solution templates. Works on benchmarks. Breaks on follow-up questions.
|
|
@@ -245,7 +255,4 @@ For the full training architecture methodology: [V3.5 Architecture Spec](https:/
|
|
| 245 |
---
|
| 246 |
|
| 247 |
*Built by [Saul Verdugo](https://huggingface.co/Verdugie) — independent ML researcher. OpusReasoning@proton.me*
|
| 248 |
-
<!-- build: 2026-03-10 19:
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
|
|
|
| 30 |
|
| 31 |
---
|
| 32 |
|
| 33 |
+
## Why "Oracle"
|
| 34 |
+
|
| 35 |
+
In computer science, an **oracle** is a theoretical construct — a black box that always returns the correct answer to a given problem, regardless of complexity. You don't ask an oracle *how* it knows. You ask it a question and it answers correctly. The concept appears across computational complexity theory (oracle machines), cryptography (random oracles), and formal verification.
|
| 36 |
+
|
| 37 |
+
STEM-Oracle borrows the name deliberately. The goal isn't a model that *searches* for answers — it's a model where the correct reasoning disposition is already in the weights. Ask it about gauge invariance and it doesn't pattern-match keywords to a textbook paragraph. It reconstructs the concept from trained understanding, meets you at your level, and corrects your mistakes along the way. The oracle metaphor captures what we're training for: a model whose *default state* is knowing how to teach, not one that needs to be prompted into it.
|
| 38 |
+
|
| 39 |
+
The "-Soul" suffix in the shared architecture name (oracle-soul) refers to the training philosophy: the model's pedagogical character lives in the weight space itself, not in a system prompt that any user can strip away. The soul survives quantization. That's the thesis.
|
| 40 |
+
|
| 41 |
+
---
|
| 42 |
+
|
| 43 |
## What Makes This Different from Math-Distilled Models
|
| 44 |
|
| 45 |
Standard STEM fine-tunes (WizardMath, MetaMath, etc.) train on solution-answer pairs. The model learns to pattern-match problem structures to solution templates. Works on benchmarks. Breaks on follow-up questions.
|
|
|
|
| 255 |
---
|
| 256 |
|
| 257 |
*Built by [Saul Verdugo](https://huggingface.co/Verdugie) — independent ML researcher. OpusReasoning@proton.me*
|
| 258 |
+
<!-- build: 2026-03-10 19:07 -->
|
|
|
|
|
|
|
|
|