Pretty interesting. I like the prompt rewriting approach they have. I think I also saw they use the ontology to have the LLM generate examples (for context) that match the user set ontology, which is nice and makes sense.
LLMS
-
LoRA adapters represent the future of efficient fine-tuning
By
–
5/6 #LoRA adapters are the future of fine-tuning for efficiency and performance
-
Fine-tuned Solar LLM Outperforms GPT-4 on Most Tasks
By
–
2/6 Fine-tuned Solar LLM outperforms #GPT-4 in over 80% of tasks. Try it yourself!
-
Open Source LLMs Outperform GPT-4 on 85% of Tasks
By
–
4/6 Fine-Tuning Index shows #OSS #LLMs outperforming GPT-4 on 85% of tested tasks
-
NotebookLM: Google’s little-known life-changing AI
By
–
This #AI tool changed my life and yet, no one knows about it Here's NotebookLM from @Google and after this tutorial, you won't be able to do without it → https://youtu.be/0b_-Y6fG0nY #ArtificialIntelligence #Google #GoogleGemini
-
50k Followers Milestone: Python, Data Science & MLOps Content
By
–
We are 50k! Thank you everyone for all the support! If you are new here, I tweet about: – Python
– Data Science
– Machine Learning
– MLOps
– Building with LLMs Follow me if you haven't already Always looking to get better & create more value for the community! -
Gen-3 Alpha opens while AI tools and jobs expand
By
–
Top stories in AI today: -Runway opens Gen-3 Alpha access
-Motorola hits the AI runway
-Visualize projects with Claude Artifacts
-New AI framework optimizes LLM usage
-6 new AI tools & 4 new AI jobs Read more: http://
therundown.ai/p/gen-3-alpha-
opens-to-the-masses
… -
Effectiveness of Plain Prompting vs Fine-Tuned Models for AI Agents
By
–
It's trendy to share models "fine-tuned for function calling", e.g. Command-R-Plus or Mixtral-8x22B. But you don't need this to make good agents Cf graph:
The count of incorrectly formatted actions is already close to 0 with plain prompting! (GPT-4o, GAIA validation run) -
Optimizing LoRA Training: Rank and Step Count Adjustments
By
–
you can also try adjusting the training settings with a different lora rank or step count. it's a tricky balance depending on the training set.
-
Reducing Concept Strength to Prevent Overfitting
By
–
Lowering concept strength to something like 0.2 when generating tends to reduce overfitting