|
|
--- |
|
|
license: mit |
|
|
inference: false |
|
|
datasets: |
|
|
- freecs/ArtificialThinkerSet |
|
|
base_model: microsoft/phi-2 |
|
|
--- |
|
|
|
|
|
# The First Open-Source Reasoning LLM |
|
|
|
|
|
**December 28, 2023** - This model was created 11 months before OpenAI's o1 release. |
|
|
|
|
|
## Historical Context |
|
|
|
|
|
In late 2023, I was experimenting with fine-tuning open-source models. Working with limited computational resources (primarily free Colab notebooks with T4 GPUs), I focused on developing novel approaches and new paradigms to significantly enhance LLM capabilities without simply scaling the number of parameters, since that would have required substantial computational resources. |
|
|
|
|
|
**Proof of timeline:** Check the [initial commit](https://huggingface.co/freecs/ArtificialThinker-Phi2/commit/8ce7acd72fb187cd3c3e76a8c0c58b8246e85d23) - December 28, 2023. |
|
|
|
|
|
## Technical Approach |
|
|
|
|
|
The model uses a custom chat template that includes a "reasoning" step before providing the output to the user: |
|
|
|
|
|
``` |
|
|
<|system|>sys_message |
|
|
<|prompt|>prompt |
|
|
<|reasoning|>reasoning |
|
|
<|response|>response<|endoftext|> |
|
|
``` |
|
|
|
|
|
To test this approach, I created the [ArtificialThinkerSet](https://huggingface.co/datasets/freecs/ArtificialThinkerSet) dataset to fine-tune Phi-2. |
|
|
|
|
|
I also wrote ["Reasoning Is All You Need"](https://freecs.org/paper.html) - a blog post explaining this approach. |
|
|
|
|
|
You can find me at [gr.bio](https://gr.bio/). |
