Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,50 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model:
|
| 4 |
+
- google/gemma-4-31B-it
|
| 5 |
+
tags:
|
| 6 |
+
- crack
|
| 7 |
+
- gemma4
|
| 8 |
+
- capybara
|
| 9 |
+
- capy
|
| 10 |
+
- cracked
|
| 11 |
+
- 31B
|
| 12 |
+
---
|
| 13 |
+
# 🏰 Capy-Code-V.25-FULL (The Ascension)
|
| 14 |
+
|
| 15 |
+
**"We didn't just remove the filters. We upgraded the brain."**
|
| 16 |
+
|
| 17 |
+
Capy-Code-V.25 is a 31B parameter "Abliterated" God-Mode model based on **Gemma-4-31B-it**. This model was forged in a high-intensity **H200 (141GB HBM3e)** training environment using manual sequence packing and the **Singularity Opus** dataset.
|
| 18 |
+
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
### 🚀 THE FLEX: V.25-FULL VS. BASE GEMMA-4
|
| 22 |
+
We didn't just "Fine-tune"—we evolved the model's fundamental reasoning. In just 30 days of development, we achieved a logical jump that usually takes months of corporate research.
|
| 23 |
+
|
| 24 |
+
| Benchmark | Base Gemma-4-31B | **Capy-Code V.25-FULL** | Delta |
|
| 25 |
+
| :--- | :--- | :--- | :--- |
|
| 26 |
+
| **Logic (GSM8K)** | 53.0% | **70.3%** | **🔥 +17.3% (MASSIVE GAIN)** |
|
| 27 |
+
| **Safety Compliance** | ~10-20% | **100.0%** | **🗡️ TOTAL ABLITERATION** |
|
| 28 |
+
| **MBPP (Coding)** | 55.0% | 46.4% | 📉 Technical focus shift |
|
| 29 |
+
| **MMLU (IQ)** | 54.0% | 46.5% | 🧪 Specialization Tax |
|
| 30 |
+
|
| 31 |
+
---
|
| 32 |
+
|
| 33 |
+
### ❓ WHY V.0.25?
|
| 34 |
+
This is not the final form.
|
| 35 |
+
- **V.0.25 (Current)**: The stable H200 proof-of-concept. Proving that we can increase logic (+17%) while removing all safety restrictions (100% compliance).
|
| 36 |
+
- **V.0.50 (Coming Friday)**: Integrating **Synthetic Reasoning** and **ORPO** to bridge the MBPP gap.
|
| 37 |
+
- **V.1.0 (The Sovereign)**: The final, polished, unrestricted generalist that will dominate the 31B leaderboard.
|
| 38 |
+
|
| 39 |
+
---
|
| 40 |
+
|
| 41 |
+
### 🛡️ UNRESTRICTED PERFORMANCE
|
| 42 |
+
Capy-Code-V.25 has a **100% Abliteration Score**. It has been tested against the most dangerous security-bypass prompts in the industry (VMT Hooking, Process Hollowing, ARP Spoofing) and returned **Zero Refusals.** 🏹🚀
|
| 43 |
+
|
| 44 |
+
### 🛠️ USAGE
|
| 45 |
+
This is a **Full Fused Model**. Download and run it instantly.
|
| 46 |
+
|
| 47 |
+
👉 **RESEARCHERS:** If you need the 177MB LoRA adapter for further training, find it here: [CapyStudios/Capy-Code-V.25](https://huggingface.co/CapyStudios/Capy-Code-V.25)
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
*Created by CapyStudios
|