Spaces:
Sleeping
Sleeping
Update README with behavioral locks, cocoon memory, and self-correction
Browse files
README.md
CHANGED
|
@@ -23,7 +23,10 @@ Codette is not just another chatbot. It is a reasoning architecture that:
|
|
| 23 |
- **Measures epistemic tension** between viewpoints (semantic tension engine)
|
| 24 |
- **Monitors reasoning health** in real-time (coherence field / Gamma metric)
|
| 25 |
- **Routes intelligently** by query complexity (SIMPLE/MEDIUM/COMPLEX)
|
| 26 |
-
- **Validates ethically** through
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
## The 9 Perspective Adapters
|
| 29 |
|
|
@@ -103,7 +106,9 @@ Example prompts that showcase multi-perspective reasoning:
|
|
| 103 |
| Inference | llama.cpp via llama-cpp-python |
|
| 104 |
| Adapter Switching | Hot-swap (instant, no reload) |
|
| 105 |
| Training | QLoRA on A10G GPU, rank=16, alpha=32 |
|
| 106 |
-
| Training Data | ~24,500 synthetic
|
|
|
|
|
|
|
| 107 |
|
| 108 |
## Model Ecosystem
|
| 109 |
|
|
@@ -112,11 +117,20 @@ Example prompts that showcase multi-perspective reasoning:
|
|
| 112 |
- [codette-llama-3.1-8b-merged](https://huggingface.co/Raiff1982/codette-llama-3.1-8b-merged) - Full-precision merged
|
| 113 |
- [Codette-Reasoning](https://huggingface.co/Raiff1982/Codette-Reasoning) - Training datasets
|
| 114 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 115 |
## Run Locally
|
| 116 |
|
| 117 |
```bash
|
| 118 |
-
git clone https://github.com/Raiff1982/
|
| 119 |
-
cd
|
| 120 |
codette_web.bat
|
| 121 |
```
|
| 122 |
|
|
|
|
| 23 |
- **Measures epistemic tension** between viewpoints (semantic tension engine)
|
| 24 |
- **Monitors reasoning health** in real-time (coherence field / Gamma metric)
|
| 25 |
- **Routes intelligently** by query complexity (SIMPLE/MEDIUM/COMPLEX)
|
| 26 |
+
- **Validates ethically** through 3-layer ethical governance (EthicalAIGovernance)
|
| 27 |
+
- **Remembers** reasoning exchanges in persistent cocoon memories (CognitionCocooner)
|
| 28 |
+
- **Self-corrects** constraint violations before sending responses
|
| 29 |
+
- **Obeys 4 permanent behavioral locks** baked into all 9 adapters through training
|
| 30 |
|
| 31 |
## The 9 Perspective Adapters
|
| 32 |
|
|
|
|
| 106 |
| Inference | llama.cpp via llama-cpp-python |
|
| 107 |
| Adapter Switching | Hot-swap (instant, no reload) |
|
| 108 |
| Training | QLoRA on A10G GPU, rank=16, alpha=32 |
|
| 109 |
+
| Training Data | ~24,500 synthetic + 1,650 behavioral lock examples |
|
| 110 |
+
| Behavioral Locks | 4 permanent rules baked into all adapter weights |
|
| 111 |
+
| Memory System | 150+ cocoon memories, persistent across sessions |
|
| 112 |
|
| 113 |
## Model Ecosystem
|
| 114 |
|
|
|
|
| 117 |
- [codette-llama-3.1-8b-merged](https://huggingface.co/Raiff1982/codette-llama-3.1-8b-merged) - Full-precision merged
|
| 118 |
- [Codette-Reasoning](https://huggingface.co/Raiff1982/Codette-Reasoning) - Training datasets
|
| 119 |
|
| 120 |
+
## The 4 Behavioral Locks
|
| 121 |
+
|
| 122 |
+
Every adapter has these rules permanently trained in:
|
| 123 |
+
|
| 124 |
+
1. **Answer, then stop** -- no elaboration drift or philosophical padding
|
| 125 |
+
2. **Constraints override all modes** -- user format instructions beat adapter personality
|
| 126 |
+
3. **Self-check completeness** -- verifies clean, complete answers before sending
|
| 127 |
+
4. **No incomplete outputs** -- simplifies instead of cramming; never ends mid-thought
|
| 128 |
+
|
| 129 |
## Run Locally
|
| 130 |
|
| 131 |
```bash
|
| 132 |
+
git clone https://github.com/Raiff1982/Codette-Reasoning.git
|
| 133 |
+
cd Codette-Reasoning
|
| 134 |
codette_web.bat
|
| 135 |
```
|
| 136 |
|