Card: soften unmeasured Semantic Attractor Clustering and SAAQ 1.5 discovery claims

#2
Files changed (1) hide show
  1. README.md +16 -16
README.md CHANGED
@@ -103,7 +103,7 @@ dataset_info:
103
 
104
  # SEMM-Latent-Telemetry
105
 
106
- > Bare-metal hardware telemetry and SNN latent space routing data for neuromorphic quantization research. This dataset documents the discovery of **Semantic Attractor Clustering** — that a Spiking Neural Network physically routes different semantic concepts (abstract language vs code syntax vs math logic) into distinct, repeatable biological pathways when L2 Normalization is applied to LLM embeddings.
107
 
108
  **Hub ID:** [`rmems/SEMM-Latent-Telemetry`](https://huggingface.co/datasets/rmems/SEMM-Latent-Telemetry)
109
  **Names:** **SEMM** = Spiking Experts Mixture Mechanism · **SAAQ** = Spiking Adaptive Activity Quantization · **Metis** = research idea that SAAQ can deepen understanding of **SNN–ANN hybrid** setups
@@ -116,7 +116,7 @@ dataset_info:
116
  - **Language(s):** English, Code
117
  - **License:** Apache 2.0 / MIT (dual-licensed)
118
  - **Repository:** [huggingface.co/datasets/rmems/SEMM-Latent-Telemetry](https://huggingface.co/datasets/rmems/SEMM-Latent-Telemetry)
119
- - **Primary routing model (first-day discovery):** [allenai/OLMoE-1B-7B-0125-Instruct](https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct) · GGUF: [allenai/OLMoE-1B-7B-0125-Instruct-GGUF](https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct-GGUF)
120
  - **SR campaign models:** see [Models Tested](#models-tested-sr-benchmark) (8 GGUF lineups + OLMoE baseline)
121
  - **Implementation:** [corinth-canal](https://github.com/Limen-Neural/corinth-canal) (SNN quantization pipeline)
122
  - **Analysis:** [Surrogate_Viz.jl](https://github.com/Spikenaut/Surrogate_Viz.jl) (symbolic regression)
@@ -136,28 +136,28 @@ When I pitched this idea, most people didn't believe the spike data conversion w
136
 
137
  ### Relationship to Spikenaut
138
 
139
- **Spikenaut** is my pure SNN model, built from scratch as a native spiking neural network. **Metis** is the research program and teacher-side probe: SAAQ applied through the OLMoE Mixture-of-Experts model to discover routing laws and adaptation equations. Those discoveries are meant to transfer into **SEMM** and into Spikenaut's native training and evolution. Metis proves the math; Spikenaut implements it natively.
140
 
141
  ---
142
 
143
- ## The Science: Semantic Attractor Clustering
144
 
145
  This dataset contains the raw bare-metal telemetry logs and latent space visualizations generated by the routing encoder. The objective is to map the physical routing of LLM embeddings (specifically from the allenai/OLMoE-1B-7B-0125-Instruct-GGUF Mixture of Experts model) as they are processed by biologically-inspired neuronal fatigue mechanics.
146
 
147
- ### The Discovery: Physical Neighborhood Mapping
148
 
149
- The primary breakthrough documented in this dataset is the organic, physical separation of semantic concepts into distinct routing bands. By applying L2 Normalization to the embeddings, the network bounds semantic pressure, forcing tokens to follow the biological **path of least resistance**.
150
 
151
- Telemetry visualizations prove that the Spike-based routing physically routes different cognitive tasks into isolated biological neighborhoods:
152
 
153
  | Abstract Language Routing (2000-Route) | Structured Logic Routing (600-800 Band) |
154
  |:---:|:---:|
155
  | ![English Logic Routing](first-day-testing-real-weights/second-test/map_olmoe_english_logic.png) | ![Rust Syntax Routing](first-day-testing-real-weights/third-test/map_olmoe_rust_syntax_logic.png) |
156
  | Abstract English logic establishes a dominant attractor basin at the **2000-index walker route** | Rust syntax and math logic collapse into the same **600-800 frequency band** |
157
 
158
- When fed abstract English logic, the network distributes energy across multiple nodes, establishing a dominant attractor basin at the **2000-index walker route**, with secondary echoes in Walkers 700 and 1450.
159
 
160
- When fed rigid mathematical statements or raw Rust syntax, the network completely abandons the 2000-route. The tokens experience mathematical pushback in abstract centers and organically collapse into the exact same **600-800 frequency band**. This demonstrates that the network physically maps highly structured logic tasks to adjacent biological neighborhoods to conserve energy.
161
 
162
  ---
163
 
@@ -169,9 +169,9 @@ The dataset documents the chronological progression from synthetic baselines to
169
  |:---:|------|-------|------------|
170
  | 1 | Synthetic Baseline | Synthetic sine wave | Verified GPU temporal loop (10,000 ticks) and basic biological fatigue |
171
  | 2 | F16 Magnitude Collapse | Real LLM embeddings (OLMoE) | Unscaled F16-to-F32 extraction caused routing collapse (single walker overwhelmed) |
172
- | 3 | Attractor Discovery | "Let's teach this MoE model..." | L2 Normalization shattered the collapse; energy settled into Walker 2000 |
173
  | 4 | Rust Syntax | `fn main() { println!(); }` | Code syntax routed to completely different biological neighborhood (600-800 band) |
174
- | 5 | Math Logic Clustering | "The derivative of a constant is mathematically zero." | Math logic routed to the same 600-800 band as Rust syntax — confirming Semantic Attractor Clustering |
175
 
176
  ---
177
 
@@ -203,7 +203,7 @@ Slug names match `sr_benchmark/{slug}/` and the `model` column in the `sr_pareto
203
  | `gemma4_26b_a4b_iq4_nl` | IQ4_NL | [google/gemma-4-26B-A4B-it](https://huggingface.co/google/gemma-4-26B-A4B-it) | e.g. [unsloth/gemma-4-26B-A4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) (UD-IQ4_NL) | math_logic, rust_syntax, english_snn |
204
  | `llama_3_2_dark_champion_q5_k_m` | Q5_K_M | [DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B) | [DavidAU/…-18.4B-GGUF](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF) | math_logic, rust_syntax, english_snn |
205
 
206
- **First-day Semantic Attractor Clustering maps** (English / Rust / math) were produced with **OLMoE** only. The multi-model table is the **symbolic regression** campaign over SAAQ features (often RE4-backed drive + prompt embeddings), not a re-run of every routing map on every model.
207
 
208
  ### Files Per Model
209
 
@@ -211,11 +211,11 @@ Each model directory contains Pareto front CSVs (`{condition}_pareto.csv`) and r
211
 
212
  ### Key Result
213
 
214
- The SAAQ 1.5 delta-Q adaptation law was discovered from this benchmark:
215
 
216
  $$\Delta q_{\text{target}} \approx 0.0573 \cdot \sqrt{\bar{f}_{\text{pop}}} + 0.496 \cdot \Delta q_{\text{prev}}$$
217
 
218
- See `SAAQ 1.5/` for the full derivation and interpretation.
219
 
220
  ---
221
 
@@ -325,7 +325,7 @@ This data behaves as "sensorimotor" stimulus for neural networks:
325
  │ ├── build_sr_pareto_all.py # Build consolidated SR table
326
  │ └── validate_dataset.py # Local interface checks (CI later)
327
  ├── SAAQ 1.0/ # Foundational L2-normalized voltage bound
328
- ├── SAAQ 1.5/ # Delta-Q adaptation law (discovered via SR)
329
  ├── CONTEXT.md # Domain glossary + multi-config decisions
330
  ├── Research notes/ # Informal research notes
331
  └── model_metadata_manifest.json # Structured project metadata
@@ -383,7 +383,7 @@ Montoya Cardenas, R. (2026). SEMM-Latent-Telemetry: Spiking Experts Mixture Mech
383
  |------|------------|
384
  | **Walker** | A pulse of electrical energy (spike) that physically explores the network to find the path of least resistance. Analogous to electrical impulses in a biological brain. |
385
  | **L2 Normalization** | Bounds semantic pressure to the unit sphere, preventing any single neuron from becoming dominant. Mimics biological brain energy distribution. |
386
- | **Semantic Attractor Clustering** | The phenomenon where an SNN physically maps different semantic concepts (abstract language vs code syntax) to distinct, repeatable biological pathways. |
387
  | **Fatigue Mechanics** | Neurons that fire too much become less responsive, preventing energy overload and enabling network adaptation. |
388
  | **SAAQ** | Spiking Adaptive Activity Quantization — the activity-aware quantization technique developed through this research (L2 voltage bound + delta-q adaptation). |
389
  | **SEMM** | Spiking Experts Mixture Mechanism — spike-driven experts / mixture routing under study (this dataset’s subject). |
 
103
 
104
  # SEMM-Latent-Telemetry
105
 
106
+ > Bare-metal hardware telemetry and SNN latent space routing data for neuromorphic quantization research. It documents early runs behind a hypothesis we call **Semantic Attractor Clustering**: that a Spiking Neural Network routes different semantic concepts (abstract language vs code syntax vs math logic) into distinct pathways when L2 Normalization is applied to LLM embeddings. This is a qualitative observation from single-prompt runs; it is not yet measured, and no clustering metric is published.
107
 
108
  **Hub ID:** [`rmems/SEMM-Latent-Telemetry`](https://huggingface.co/datasets/rmems/SEMM-Latent-Telemetry)
109
  **Names:** **SEMM** = Spiking Experts Mixture Mechanism · **SAAQ** = Spiking Adaptive Activity Quantization · **Metis** = research idea that SAAQ can deepen understanding of **SNN–ANN hybrid** setups
 
116
  - **Language(s):** English, Code
117
  - **License:** Apache 2.0 / MIT (dual-licensed)
118
  - **Repository:** [huggingface.co/datasets/rmems/SEMM-Latent-Telemetry](https://huggingface.co/datasets/rmems/SEMM-Latent-Telemetry)
119
+ - **Primary routing model (first-day exploratory runs):** [allenai/OLMoE-1B-7B-0125-Instruct](https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct) · GGUF: [allenai/OLMoE-1B-7B-0125-Instruct-GGUF](https://huggingface.co/allenai/OLMoE-1B-7B-0125-Instruct-GGUF)
120
  - **SR campaign models:** see [Models Tested](#models-tested-sr-benchmark) (8 GGUF lineups + OLMoE baseline)
121
  - **Implementation:** [corinth-canal](https://github.com/Limen-Neural/corinth-canal) (SNN quantization pipeline)
122
  - **Analysis:** [Surrogate_Viz.jl](https://github.com/Spikenaut/Surrogate_Viz.jl) (symbolic regression)
 
136
 
137
  ### Relationship to Spikenaut
138
 
139
+ **Spikenaut** is my pure SNN model, built from scratch as a native spiking neural network. **Metis** is the research program and teacher-side probe: SAAQ applied through the OLMoE Mixture-of-Experts model to search for candidate routing laws and adaptation equations. Candidates that hold up under testing are meant to transfer into **SEMM** and into Spikenaut's native training and evolution. Metis tests the math; Spikenaut implements it natively.
140
 
141
  ---
142
 
143
+ ## The Hypothesis: Semantic Attractor Clustering
144
 
145
  This dataset contains the raw bare-metal telemetry logs and latent space visualizations generated by the routing encoder. The objective is to map the physical routing of LLM embeddings (specifically from the allenai/OLMoE-1B-7B-0125-Instruct-GGUF Mixture of Experts model) as they are processed by biologically-inspired neuronal fatigue mechanics.
146
 
147
+ ### Qualitative Observation: Routing Bands
148
 
149
+ We hypothesize that semantic concepts separate into distinct routing bands. Applying L2 Normalization to the embeddings bounds semantic pressure, so we expect tokens to follow the **path of least resistance**.
150
 
151
+ The visualizations below qualitatively suggest that spike-based routing sends different prompt types to different walker ranges. Each condition is a single prompt and a single run, with no repeats or controls:
152
 
153
  | Abstract Language Routing (2000-Route) | Structured Logic Routing (600-800 Band) |
154
  |:---:|:---:|
155
  | ![English Logic Routing](first-day-testing-real-weights/second-test/map_olmoe_english_logic.png) | ![Rust Syntax Routing](first-day-testing-real-weights/third-test/map_olmoe_rust_syntax_logic.png) |
156
  | Abstract English logic establishes a dominant attractor basin at the **2000-index walker route** | Rust syntax and math logic collapse into the same **600-800 frequency band** |
157
 
158
+ For the abstract English prompt, the map appears to show a dominant route near the **2000-index walker**, with secondary activity near Walkers 700 and 1450. This is read from the image only; no per-tick log is published for this run.
159
 
160
+ For the Rust syntax prompt and the math prompt, routing concentrates in the **600-800 walker band**. From the published per-tick logs, the best walker fell in 600-800 on 8,566 of 10,000 ticks for Rust (`third-test/map_olmoe_rust_syntax_logic.txt`, mode walker 747 on 8,089 ticks) and on 9,123 of 10,000 ticks for math (`fourth-test/telemetry_olmoe_math_logic.txt`, mode walker 746 on 8,793 ticks). Whether this reflects task structure (or energy conservation) is untested: there is one prompt per condition, no repeated runs or baselines, and no clustering or separation metric is published.
161
 
162
  ---
163
 
 
169
  |:---:|------|-------|------------|
170
  | 1 | Synthetic Baseline | Synthetic sine wave | Verified GPU temporal loop (10,000 ticks) and basic biological fatigue |
171
  | 2 | F16 Magnitude Collapse | Real LLM embeddings (OLMoE) | Unscaled F16-to-F32 extraction caused routing collapse (single walker overwhelmed) |
172
+ | 3 | L2 Normalization (English) | "Let's teach this MoE model..." | L2 Normalization removed the collapse; the map appears to show energy settling near Walker 2000 (image only) |
173
  | 4 | Rust Syntax | `fn main() { println!(); }` | Code syntax routed to completely different biological neighborhood (600-800 band) |
174
+ | 5 | Math Logic | "The derivative of a constant is mathematically zero." | Best walker in 600-800 on 9,123 of 10,000 ticks, the same band as Rust syntax; consistent with, but not a test of, the Semantic Attractor Clustering hypothesis |
175
 
176
  ---
177
 
 
203
  | `gemma4_26b_a4b_iq4_nl` | IQ4_NL | [google/gemma-4-26B-A4B-it](https://huggingface.co/google/gemma-4-26B-A4B-it) | e.g. [unsloth/gemma-4-26B-A4B-it-GGUF](https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF) (UD-IQ4_NL) | math_logic, rust_syntax, english_snn |
204
  | `llama_3_2_dark_champion_q5_k_m` | Q5_K_M | [DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B) | [DavidAU/…-18.4B-GGUF](https://huggingface.co/DavidAU/Llama-3.2-8X3B-MOE-Dark-Champion-Instruct-uncensored-abliterated-18.4B-GGUF) | math_logic, rust_syntax, english_snn |
205
 
206
+ **First-day routing maps** (English / Rust / math) were produced with **OLMoE** only. The multi-model table is the **symbolic regression** campaign over SAAQ features (often RE4-backed drive + prompt embeddings), not a re-run of every routing map on every model.
207
 
208
  ### Files Per Model
209
 
 
211
 
212
  ### Key Result
213
 
214
+ A candidate SAAQ 1.5 delta-Q adaptation law was proposed from this campaign (see `SAAQ 1.5/`). Its fit loss is not in the published Pareto files (`sr_pareto_all`, `SAAQ 1.5/hall_of_fame.csv`), so treat it as a proposed law, not a validated one:
215
 
216
  $$\Delta q_{\text{target}} \approx 0.0573 \cdot \sqrt{\bar{f}_{\text{pop}}} + 0.496 \cdot \Delta q_{\text{prev}}$$
217
 
218
+ See `SAAQ 1.5/` for the raw expression and interpretation.
219
 
220
  ---
221
 
 
325
  │ ├── build_sr_pareto_all.py # Build consolidated SR table
326
  │ └── validate_dataset.py # Local interface checks (CI later)
327
  ├── SAAQ 1.0/ # Foundational L2-normalized voltage bound
328
+ ├── SAAQ 1.5/ # Delta-Q adaptation law (proposed via SR)
329
  ├── CONTEXT.md # Domain glossary + multi-config decisions
330
  ├── Research notes/ # Informal research notes
331
  └── model_metadata_manifest.json # Structured project metadata
 
383
  |------|------------|
384
  | **Walker** | A pulse of electrical energy (spike) that physically explores the network to find the path of least resistance. Analogous to electrical impulses in a biological brain. |
385
  | **L2 Normalization** | Bounds semantic pressure to the unit sphere, preventing any single neuron from becoming dominant. Mimics biological brain energy distribution. |
386
+ | **Semantic Attractor Clustering** | Working hypothesis that an SNN maps different semantic concepts (abstract language vs code syntax) to distinct, repeatable pathways. Qualitatively observed in single runs; not yet measured. |
387
  | **Fatigue Mechanics** | Neurons that fire too much become less responsive, preventing energy overload and enabling network adaptation. |
388
  | **SAAQ** | Spiking Adaptive Activity Quantization — the activity-aware quantization technique developed through this research (L2 voltage bound + delta-q adaptation). |
389
  | **SEMM** | Spiking Experts Mixture Mechanism — spike-driven experts / mixture routing under study (this dataset’s subject). |