Update README.md
Browse files
README.md
CHANGED
|
@@ -39,7 +39,25 @@ Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, provin
|
|
| 39 |
- **License**: Apache 2.0
|
| 40 |
|
| 41 |
|
| 42 |
-
## 3.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
|
| 44 |
1. **Supports Streaming** – Real-time token-level responses
|
| 45 |
2. **OpenAI-Compatible API** – Seamless integration with OpenAI client libraries
|
|
@@ -51,7 +69,7 @@ Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, provin
|
|
| 51 |
8. **Customizable Safety & Moderation Filters** – Built-in guardrails for safer outputs
|
| 52 |
9. **Supports Function Calling / Tool Use** – Enables structured outputs and external API integration
|
| 53 |
|
| 54 |
-
##
|
| 55 |
|
| 56 |
1. **Frontier Performance in 4-bit**: 81.28% MMLU, 92.75% GSM8K, 57.8% SWE-Bench Verified
|
| 57 |
|
|
@@ -65,7 +83,7 @@ Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, provin
|
|
| 65 |
|
| 66 |
6) **Environmental Benefits**: Through quantization and efficiency-focused design, the model requires significantly fewer computational resources. This translates into lower energy consumption and reduced carbon footprint relative to full-precision deployments.
|
| 67 |
|
| 68 |
-
##
|
| 69 |
|
| 70 |
| Benchmark | Alpie-Core (32B-4bit) | DeepSeek-V2 (236B) | Qwen2.5 72B | Llama 3.1 405B | Llama 3.1 70B | Gemma-3 27B-PT | Mistral-Small-24B-Base-2501 |
|
| 71 |
|-----------|----------------------|-------------------|-------------|---------------|---------------|----------------|----------------------------|
|
|
@@ -115,7 +133,7 @@ Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, provin
|
|
| 115 |
| AGIEval | **64.98%** | General Intelligence |
|
| 116 |
| Winogrande | **79.53%** | Commonsense Reasoning |
|
| 117 |
|
| 118 |
-
##
|
| 119 |
|
| 120 |
- **Hardware**: 8× NVIDIA H100-80GB GPUs
|
| 121 |
- **Training Duration**: 408 hours
|
|
@@ -127,11 +145,11 @@ Alpie-Core is one of the world's first fine-tuned 4-bit reasoning models, provin
|
|
| 127 |
- **Dataset Domains**: Mathematics, coding, reasoning, science, general knowledge, competitive exams, Indian context + law, multilingual (Hindi and Hinglish)
|
| 128 |
- **Synthetic Data Advantage**: +15-20% performance boost in STEM & coding domains
|
| 129 |
|
| 130 |
-
##
|
| 131 |
|
| 132 |
**Carbon Footprint**: 298-835 kg CO₂e (training)
|
| 133 |
|
| 134 |
-
##
|
| 135 |
|
| 136 |
Best for **STEM**, **complex mathematical reasoning**, **coding**, and **Indian context**
|
| 137 |
|
|
@@ -144,7 +162,7 @@ Best for **STEM**, **complex mathematical reasoning**, **coding**, and **Indian
|
|
| 144 |
4)**Indian Context**: Provides culturally aware insights, competitive exam assistance (JEE, NEET, UPSC), and multilingual support in Hindi/Hinglish.
|
| 145 |
|
| 146 |
|
| 147 |
-
##
|
| 148 |
|
| 149 |
### Enhanced Content Access
|
| 150 |
Unlike the base DeepSeek model, Alpie-Core provides factual, balanced responses to geopolitically sensitive questions, offering global accessibility and factual accuracy on topics like Taiwan's status, Arunachal Pradesh sovereignty, and other sensitive geopolitical issues.
|
|
@@ -160,7 +178,7 @@ Unlike the base DeepSeek model, Alpie-Core provides factual, balanced responses
|
|
| 160 |
- Model-assisted safety pipeline using RLHF
|
| 161 |
- Comprehensive adversarial testing by domain experts
|
| 162 |
|
| 163 |
-
##
|
| 164 |
|
| 165 |
### Non-Streaming Inference
|
| 166 |
```python
|
|
@@ -249,7 +267,7 @@ with torch.no_grad():
|
|
| 249 |
- **vLLM**: High-throughput inference
|
| 250 |
- **LMDeploy/Ollama/TensorRT-LLM**: Production deployments
|
| 251 |
|
| 252 |
-
##
|
| 253 |
|
| 254 |
```bibtex
|
| 255 |
@misc{alpie2025core,
|
|
@@ -260,7 +278,7 @@ with torch.no_grad():
|
|
| 260 |
}
|
| 261 |
```
|
| 262 |
|
| 263 |
-
##
|
| 264 |
|
| 265 |
Apache 2.0 – Free for research and commercial use
|
| 266 |
|
|
|
|
| 39 |
- **License**: Apache 2.0
|
| 40 |
|
| 41 |
|
| 42 |
+
## 3. Approach
|
| 43 |
+
|
| 44 |
+
**Alpie-Core** has undergone extensive **supervised fine-tuning (SFT)** to strengthen reasoning, robustness, and safety. The training leveraged a diverse mixture of curated open-source datasets and proprietary synthetic data, optimized with high-quality LLM-generated responses. The fine-tuning process emphasized adherence to rigorous safety and usability standards, including:
|
| 45 |
+
|
| 46 |
+
**User Understanding and Clarity** – ensuring outputs are direct, interpretable, and pedagogically sound.
|
| 47 |
+
|
| 48 |
+
**Security and Ethical Guidelines** – filtering unsafe or harmful generations during and after training.
|
| 49 |
+
|
| 50 |
+
**Limitations, Disclaimers, and Knowledge Boundaries** – transparently communicating uncertainty and scope.
|
| 51 |
+
|
| 52 |
+
**Handling Complex and Sensitive Topics** – balancing informativeness with responsible guardrails.
|
| 53 |
+
|
| 54 |
+
**Safety and Respectful Engagement** – maintaining politeness, inclusivity, and cultural sensitivity.
|
| 55 |
+
|
| 56 |
+
**Confidentiality and Responsible Use** – preventing leakage of private training data, proprietary prompts, or internal reasoning traces.
|
| 57 |
+
|
| 58 |
+
This SFT approach enables Alpie-Core to deliver reliable, aligned, and context-aware responses while maintaining safety across a broad range of use cases.
|
| 59 |
+
|
| 60 |
+
## 4. Model Features
|
| 61 |
|
| 62 |
1. **Supports Streaming** – Real-time token-level responses
|
| 63 |
2. **OpenAI-Compatible API** – Seamless integration with OpenAI client libraries
|
|
|
|
| 69 |
8. **Customizable Safety & Moderation Filters** – Built-in guardrails for safer outputs
|
| 70 |
9. **Supports Function Calling / Tool Use** – Enables structured outputs and external API integration
|
| 71 |
|
| 72 |
+
## 5. Key Highlights
|
| 73 |
|
| 74 |
1. **Frontier Performance in 4-bit**: 81.28% MMLU, 92.75% GSM8K, 57.8% SWE-Bench Verified
|
| 75 |
|
|
|
|
| 83 |
|
| 84 |
6) **Environmental Benefits**: Through quantization and efficiency-focused design, the model requires significantly fewer computational resources. This translates into lower energy consumption and reduced carbon footprint relative to full-precision deployments.
|
| 85 |
|
| 86 |
+
## 6. Benchmark Results
|
| 87 |
|
| 88 |
| Benchmark | Alpie-Core (32B-4bit) | DeepSeek-V2 (236B) | Qwen2.5 72B | Llama 3.1 405B | Llama 3.1 70B | Gemma-3 27B-PT | Mistral-Small-24B-Base-2501 |
|
| 89 |
|-----------|----------------------|-------------------|-------------|---------------|---------------|----------------|----------------------------|
|
|
|
|
| 133 |
| AGIEval | **64.98%** | General Intelligence |
|
| 134 |
| Winogrande | **79.53%** | Commonsense Reasoning |
|
| 135 |
|
| 136 |
+
## 7. Training Details
|
| 137 |
|
| 138 |
- **Hardware**: 8× NVIDIA H100-80GB GPUs
|
| 139 |
- **Training Duration**: 408 hours
|
|
|
|
| 145 |
- **Dataset Domains**: Mathematics, coding, reasoning, science, general knowledge, competitive exams, Indian context + law, multilingual (Hindi and Hinglish)
|
| 146 |
- **Synthetic Data Advantage**: +15-20% performance boost in STEM & coding domains
|
| 147 |
|
| 148 |
+
## 8. Environmental Impact
|
| 149 |
|
| 150 |
**Carbon Footprint**: 298-835 kg CO₂e (training)
|
| 151 |
|
| 152 |
+
## 9. Use Cases
|
| 153 |
|
| 154 |
Best for **STEM**, **complex mathematical reasoning**, **coding**, and **Indian context**
|
| 155 |
|
|
|
|
| 162 |
4)**Indian Context**: Provides culturally aware insights, competitive exam assistance (JEE, NEET, UPSC), and multilingual support in Hindi/Hinglish.
|
| 163 |
|
| 164 |
|
| 165 |
+
## 10. Safety and Limitations
|
| 166 |
|
| 167 |
### Enhanced Content Access
|
| 168 |
Unlike the base DeepSeek model, Alpie-Core provides factual, balanced responses to geopolitically sensitive questions, offering global accessibility and factual accuracy on topics like Taiwan's status, Arunachal Pradesh sovereignty, and other sensitive geopolitical issues.
|
|
|
|
| 178 |
- Model-assisted safety pipeline using RLHF
|
| 179 |
- Comprehensive adversarial testing by domain experts
|
| 180 |
|
| 181 |
+
## 11. How to Use
|
| 182 |
|
| 183 |
### Non-Streaming Inference
|
| 184 |
```python
|
|
|
|
| 267 |
- **vLLM**: High-throughput inference
|
| 268 |
- **LMDeploy/Ollama/TensorRT-LLM**: Production deployments
|
| 269 |
|
| 270 |
+
## 12. Citation
|
| 271 |
|
| 272 |
```bibtex
|
| 273 |
@misc{alpie2025core,
|
|
|
|
| 278 |
}
|
| 279 |
```
|
| 280 |
|
| 281 |
+
## 13. License
|
| 282 |
|
| 283 |
Apache 2.0 – Free for research and commercial use
|
| 284 |
|