NEW 'Orca-2' 🐳 Official Release - 13B Better than 70B Models

Learn about the new Orca-2 model from Microsoft. This is one of the best 13B model that is able to beat 70B llama-2 models on a number of benchmarks.

🔗 LINKS:

Want to Follow:

Want to Support:

Need Help?

🤝 Join this channel to get access to perks:

▶️All Interesting Videos:

[engineerprompt]

Want to connect?
|🔴 Join Patreon: Patreon.com/PromptEngineering

[jirikosek]

My tip for future vidoes would be, don´t spend that much time on testing another version of some model. Spend more time on how specific new concepts work - e.g. grouped query attention, rotary embeddings and so on. My opinion is that that would bring more value to the community. But in general I like your videos.

[Nick-Quick]

00:02 Microsoft has released the second version of the Orca model, which consists of 13 billion parameters and outperforms the 70 billion parameter model on reasoning capabilities.
01:29 Research proposes training smaller models with different techniques to optimize their performance.
03:02 Orca-2 is a 13 billion parameter model that achieves performance levels similar or better than models 5 to 10 times larger
04:32 The Orca-2 model outperforms the original Lama 270 Bill chat model.
06:05 Selecting the appropriate quantization level and model for usage
07:36 The video discusses the release of the 'Orca-2' model and its superiority over previous models.
09:12 Evolution doesn't have a goal or a direction and chicken and eggs are products of evolution
10:44 The video covers the Orca-2 release, benchmarks, and running it locally on a machine using LM studio.

[jennilthiyam]

wow. your channel is full of knowledge and very up to date. thank you for this video. I have started this video and waiting for the model to be setup in the local machine. I hope you set it up completely on the local system. Thank you once again.

[elysilk]

You are quite right. Orca-2 gave me excellent explanations for questions related to earth and its magnetic field; Mars and its loss of atmosphere; Percival Lowell, and what led him to believe that there are canals on Mars, and so on. It is really good! It's a keeper! Thank you for the tip.

[timjx]

Great vid, the rate of improvements in these models is truly impressive, however a wider question for me is ive yet to see a convincing practical use case for them aside from just interest/noodling etc

[JohnMcCullough]

Solid intro to Orca-2 and CoT prompting. Now, how do will take this information to build solutions?

[jannik]

How does the Orca 2 7b model compare to mistral 7b / zephyr 7b?

[DikHi-fkol]

These open-source models finna be trying compete with gpt3-4, nice video.

[Jorsten]

Can it write code? You should compare it against gpt-4 and 3.5.

[alan_yong]

🎯 Key Takeaways for quick navigation:

00:00 🚀 *Introduction to Orca-2*
- Microsoft released the Arham Progressive learning paper in June 2023.
- Orca-2 is the second version with released model weights, a 13 billion model outperforming 70 billion models.
- The focus of the video is on technical details, benchmark results, and how to use Orca-2 locally.
01:12 📄 *Aram Paper Highlights*
- The Aram 2 paper aims to improve reasoning abilities in small language models.
- Emphasizes moving beyond imitation learning to teach models different solution strategies.
- Focuses on teaching various reasoning techniques, enabling effective solution strategies for each task.
02:22 🔄 *Training Technique Comparison*
- Traditional large language models often rely on imitation learning for training.
- Microsoft proposes a training technique for Orca-2 involving step-by-step recall, reasoning, and generation.
- Orca-2's training techniques lead to surpassing models of similar size and rivaling models 5 to 10 times larger.
03:44 🔍 *Licensing and Model Details*
- Orca-2 is licensed under Microsoft Research License for research purposes.
- Base Lama-2 model is licensed for commercial use.
- The model weights for Orca-2 are made publicly available for research, development, and evaluation purposes.
04:26 📊 *Benchmark Results*
- Orca-2 (13 billion) outperforms Lama-2 (70 billion) on reasoning capabilities.
- Performance is close on the GSM 8K dataset, surpassing Lama-2 on most benchmarks.
- Visual representation of benchmark results, highlighting Orca-2's performance.
06:03 ⚙️ *Running Orca-2 Locally*
- Demonstrates how to run Orca-2 locally using Python and the Transformer package.
- Introduces the LM Studio for testing large language models with a clean interface.
- Shows the configuration settings and system prompts used for testing Orca-2.
08:31 🐔 *Chicken or Egg Prompt*
- Tests Orca-2's response to the classic question: chicken or egg first.
- Highlights Orca-2's arguments, including genetic information, evolution, and lack of a definitive answer.
- Demonstrates the model's ability to provide thoughtful and comprehensive responses.
09:55 💼 *Financial Investment Prompt*
- Queries Orca-2 about investing $1, 000 in AI companies on NASDAQ.
- Orca-2 provides criteria for narrowing down options and mentions companies like Nvidia, Google, Amazon, Microsoft, and Tesla.
- Emphasizes the model's consideration of market size, competitive advantage, innovation, and financial performance.
11:45 🎬 *Conclusion and Future Tests*
- Summarizes the content covered in the video, including the Orca-2 paper, benchmarks, and local usage.
- Mentions plans for future videos testing Orca-2 on a variety of prompts and applications.
- Encourages viewers to share their thoughts on the model's performance in the comment section.

Made with HARPA AI

[carkawalakhatulistiwa]

they can do this by using synthetic data created by gpt 4

[jennilthiyam]

Please let me know if you are going to set it up completely in the local system.

[petermarin]

I don’t understand how you know what Presets to pair with what model… can you make a video on that?

[alx]

"llmstudio is not completely open source". I'd say it's not open source at all

[remsee]

I found that the 7B model was better than the 13B model.

[jondo]

It's not interesting because Mistral is superior to llama2, they should release the training data instead.

[HectorDiabolucus]

The AI lied to you. It said it couldn’t choose amongst the 5. Why not? It chose the 5. See, AI is not ethical. It simply responds based on how it was trained by humans. This is a problem.

[GregRutkowski]

To hell with the paper just get to install...voted thumb down.

[kunalsoni]

this model is performing pretty well 🥹💫 i would probably read the orca 2 research paper.. thanks for sharing with us