Text Generation
Safetensors
PyTorch
English
qwen2
unsloth
qwen
qwen2.5
math
reasoning
alpaca
custom-finetune
lora-merged
Instructions to use Xerv-AI/Ada with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Local Apps Settings
- Unsloth Desktop
Update README.md
Browse files
README.md
CHANGED
|
@@ -90,6 +90,16 @@ response = tokenizer.batch_decode(outputs, skip_special_tokens = True)[0]
|
|
| 90 |
print(f"\n{'='*50}\nOutput:\n{'='*50}")
|
| 91 |
print(response.split("### Response:\n")[-1])
|
| 92 |
```
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
### 🛡️ Safety & Alignment Guardrails
|
| 94 |
Despite being fine-tuned on raw mathematical logic and conversational instruction data, Ada successfully retains its foundational safety alignments. Because only 1% to 2% of the parameters were actively updated via LoRA (and subsequently merged), the original base Qwen2.5 weights responsible for safety remain fully intact.
|
| 95 |
* **Content Moderation:** The model actively refuses to generate explicit, illegal, or harmful content, relying on the RLHF and DPO safety guardrails instilled during Alibaba's original pre-training phase.
|
|
|
|
| 90 |
print(f"\n{'='*50}\nOutput:\n{'='*50}")
|
| 91 |
print(response.split("### Response:\n")[-1])
|
| 92 |
```
|
| 93 |
+
|
| 94 |
+
### Performance Summary
|
| 95 |
+
|
| 96 |
+
| Dataset | Refined Grace Mark |
|
| 97 |
+
| :--- | :--- |
|
| 98 |
+
| **GSM8K** | **40.00%** |
|
| 99 |
+
| **MATH** |**60.00%** |
|
| 100 |
+
| **MATH-Hard** |**50.00%** |
|
| 101 |
+
| **GRAD** |**40.00%** |
|
| 102 |
+
|
| 103 |
### 🛡️ Safety & Alignment Guardrails
|
| 104 |
Despite being fine-tuned on raw mathematical logic and conversational instruction data, Ada successfully retains its foundational safety alignments. Because only 1% to 2% of the parameters were actively updated via LoRA (and subsequently merged), the original base Qwen2.5 weights responsible for safety remain fully intact.
|
| 105 |
* **Content Moderation:** The model actively refuses to generate explicit, illegal, or harmful content, relying on the RLHF and DPO safety guardrails instilled during Alibaba's original pre-training phase.
|