One thing I realized you did differently Mike, you added the [INST] Instruction [/INST] in the text string of your dataset. I was under the impression that MLX automatically adds the special token based on the original model, so you don't have to do that. But I may be
@skirano
-
LoRA vs Quantization: Optimizing AI Model Training Resources
By
–
Yes! I could have done Lora as well but I opted for the quantized version to reduce resources. If I had more than 64gb of RAM I feel Lora would have been fine.
-
MLX Framework for Apple Silicon Machine Learning
By
–
Yeah I used MLX, it’s amazing! You find all what you need here. It’s designed for Apple Silicon https://
github.com/ml-explore/mlx
-examples
… -
Reddit Instruct Dataset Released for AI Research
By
–
Huge thanks to @teknium for their work and to @Euclaise_ for providing the dataset! https://
huggingface.co/datasets/eucla
ise/reddit-instruct-curated
… -
Local Model Fine-tuning Tutorial and Data Formatting Guide
By
–
If you are interested in learning more about fine-tuning models locally, consider following me. I will post a tutorial on how I did this fine-tuning, how I formatted the data, etc.
-
NousHermes-Mixtral-8x7B-Reddit: Fine-tuned LLM Model Released
By
–
Introducing NousHermes-Mixtral-8x7B-Reddit A fine-tune of the legendary model by @teknium on 10k Reddit threads. Including subreddits like:
– AskAcademia
– AskComputerScience
– AskEconomics
– ELI5 And many more! Optimized for MLX. -
Model Trained Locally on M3 Max Over 9000 Iterations
By
–
This model was trained locally on M3 Max over 9k iterations. Link to the model
-
ChatGPT Now Offers Dual Response Options via Direct RLHF
By
–
Wow, my ChatGPT now provides me with two options to choose from for the answer, and they were generated at the same time. OpenAI is doing RLHF directly in the chat now!
-
Customizing prompts with your own instructions
By
–
You can also rewrite some parts of this prompt into your custom instructions and change them to what you want.
-
DesignerGPT Hits 110K Users in First Month Post-Launch
By
–
According to @replit deployments, DesignerGPT had 110,449 unique visitors last month with an average of 16k new ones each week. These are pretty bonkers numbers for an app that just launched a month ago. Top country: USA, understandable since GPTs are a premium-only feature.