117 lines
4.0 KiB
Markdown
117 lines
4.0 KiB
Markdown
|
---
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
- fr
|
|||
|
|
- de
|
|||
|
|
- it
|
|||
|
|
- es
|
|||
|
|
pipeline_tag: text-generation
|
|||
|
|
tags:
|
|||
|
|
- cygnisai
|
|||
|
|
- llama-3.1
|
|||
|
|
- model-merge
|
|||
|
|
- unsloth
|
|||
|
|
- enterprise-ready
|
|||
|
|
- finetuned
|
|||
|
|
- safety-aligned
|
|||
|
|
license: apache-2.0
|
|||
|
|
datasets:
|
|||
|
|
- cygnisai/Cygnis-Alpha-2-Instruct-Mix
|
|||
|
|
base_model:
|
|||
|
|
- unsloth/Meta-Llama-3.1-8B-bnb-4bit
|
|||
|
|
library_name: transformers
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
# Cygnis-Alpha-2-8B-v0.2 (Merge Version)
|
|||
|
|
|
|||
|
|
<div align="center" style="background:#06090f; border-radius:14px; border:1px solid #0f1e30; overflow:hidden; margin-bottom:20px;">
|
|||
|
|
<img src="https://huggingface.co/cygnisai/Cygnis-Alpha-2-8B-v0.1/resolve/main/Cygnis-alpha-2.png" width="100%" style="display:block;">
|
|||
|
|
</div>
|
|||
|
|
|
|||
|
|
**Cygnis-Alpha-2-8B-v0.2** is a next-generation large language model developed by **CygnisAI**. This version is the result of a strategic **Model Merge**, engineered to surpass the native reasoning capabilities of Llama 3.1 8B while maintaining optimal computational efficiency for enterprise-scale deployments.
|
|||
|
|
|
|||
|
|
## Architecture & Development
|
|||
|
|
|
|||
|
|
The model is built upon the **Llama 3.1** architecture (Auto-regressive Transformer). The CygnisAI team leveraged **Unsloth** for accelerated **Supervised Fine-Tuning (SFT)** and applied advanced **Model Merging** techniques to sharpen technical and professional response accuracy.
|
|||
|
|
|
|||
|
|
| Feature | Specification |
|
|||
|
|
| :--- | :--- |
|
|||
|
|
| **Developer** | [CygnisAI](https://huggingface.co/cygnisai) |
|
|||
|
|
| **Base Model** | Meta Llama 3.1 8B (Finetuned via Unsloth) |
|
|||
|
|
| **Context Length** | 128k tokens |
|
|||
|
|
| **Optimization** | Multilingual (Primary focus: EN/FR) |
|
|||
|
|
| **Knowledge Cutoff** | December 2023 |
|
|||
|
|
| **License** | Apache 2.0 |
|
|||
|
|
|
|||
|
|
## Performance & Benchmarks
|
|||
|
|
|
|||
|
|
Cygnis v0.2 has been rigorously tested against internal benchmarks to ensure production-grade stability. It shows significant improvements in **Instruction Following** for complex, multi-step tasks compared to the v0.1 release.
|
|||
|
|
|
|||
|
|
* **MMLU (Reasoning)**: ~69.8%
|
|||
|
|
* **IFEval (Strict Instruction)**: ~81.2%
|
|||
|
|
* **HumanEval (Coding)**: ~72.9%
|
|||
|
|
|
|||
|
|
## Quick Start
|
|||
|
|
|
|||
|
|
### Via Transformers (Pipeline)
|
|||
|
|
```python
|
|||
|
|
import transformers
|
|||
|
|
import torch
|
|||
|
|
|
|||
|
|
model_id = "cygnisai/Cygnis-Alpha-2-8B-v0.2"
|
|||
|
|
|
|||
|
|
pipeline = transformers.pipeline(
|
|||
|
|
"text-generation",
|
|||
|
|
model=model_id,
|
|||
|
|
model_kwargs={"torch_dtype": torch.bfloat16},
|
|||
|
|
device_map="auto"
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{"role": "system", "content": "You are Cygnis, the official enterprise AI assistant."},
|
|||
|
|
{"role": "user", "content": "Analyze the scalability requirements for our LLM infrastructure."}
|
|||
|
|
]
|
|||
|
|
|
|||
|
|
print(pipeline(messages, max_new_tokens=256))
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### Official Prompt Format
|
|||
|
|
The model follows the standard Llama 3.1 chat template:
|
|||
|
|
```text
|
|||
|
|
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
|
|||
|
|
|
|||
|
|
{system_prompt}<|eot_id|><|start_header_id|>user<|end_header_id|>
|
|||
|
|
|
|||
|
|
{user_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Responsibility & Safety
|
|||
|
|
|
|||
|
|
In alignment with **CygnisAI’s** commitment to responsible AI:
|
|||
|
|
1. **Guardrails**: The model integrates filters to mitigate malicious content, disinformation, and unintended biases.
|
|||
|
|
2. **Transparency**: While highly capable, Cygnis should not be used without human oversight for critical decision-making (Medical, Legal, Financial).
|
|||
|
|
3. **License**: This model is released under the **Apache License 2.0**, offering maximum flexibility for commercial and private use.
|
|||
|
|
|
|||
|
|
## Legal Notice
|
|||
|
|
This model was finetuned using the **Unsloth** library. While the underlying architecture is based on Llama 3.1, the specific weights of Cygnis-Alpha-2-8B-v0.2 are provided under the Apache 2.0 terms.
|
|||
|
|
*Built with Llama.*
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Citation
|
|||
|
|
|
|||
|
|
```bibtex
|
|||
|
|
@misc{cygnis_alpha_2_v0.1,
|
|||
|
|
author = {Simonc-44},
|
|||
|
|
title = {The Carbon Footprint of Machine Learning Training Will Plateau, Then Shrink},
|
|||
|
|
year = {2026},
|
|||
|
|
publisher = {Hugging Face},
|
|||
|
|
url = {https://huggingface.co/Simonc-44/Cygnis-Alpha-2-8B-v0.1}
|
|||
|
|
eprint={2204.05149},
|
|||
|
|
archivePrefix={arXiv},
|
|||
|
|
}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
**Professional Contact & Support**: [cygnisai/contact](https://huggingface.co/cygnisai)
|
|||
|
|
*Driven by innovation and corporate-grade rigor at CygnisAI.*
|