Training Bixtral, which is expensive + slow so mistakes are costly. The run finally finished, but then we got a sigkill -9 error (I believe it's a memory issue). I thought I lost the model, but I went into the output directory anyway, and luckily, it was there.
@mattshumer_
-
Bixtral Fine-Tuned Model Approaches Claude Opus Performance
By
–
Bixtral is a beast. A well-done fine-tune doesn’t feel far off from Opus.
-
Instruct Prompt to Base Model Converter Notebook
By
–
If you'd like to try it or contribute, check out the Github repo. Use the `Instruct_Prompt_>_Base_Model_Prompt_Converter.ipynb` notebook.
-
Python Interpreter Prompt Format for Mixtral 8x22B
By
–
This system uses the Python interpreter-style prompt format I describe in the article. It's flexible, and works well for most tasks. If you're building w/ Mixtral 8x22B, you NEED to try this.
-
Open-source tool converts chat prompts for Mixtral base model
By
–
Introducing `instruct-to-base-prompt-converter` ✍️
— Matt Shumer (@mattshumer_) 12 avril 2024
Prompting base models like Mixtral 8x22B is insanely hard.
Now, AI can do it for you!
Uses Claude 3 Opus to automatically convert a chat prompt to a prompt that works with the Mixtral-8x22B base model!
And it's open-source! pic.twitter.com/rKAgkXxidnIntroducing `instruct-to-base-prompt-converter` Prompting base models like Mixtral 8x22B is insanely hard. Now, AI can do it for you! Uses Claude 3 Opus to automatically convert a chat prompt to a prompt that works with the Mixtral-8x22B base model! And it's open-source!
-
Claude 3 Prompt Engineering for Base Models Guide
By
–
This repo was inspired by this article I wrote, explaining how to prompt Mixtral 8x22B. Turns out, Claude 3 is quite good at prompting base models, if you set it up correctly.
-
Prompting Evolution: From GPT-2/3 Difficulty to Modern Ease
By
–
I come from the GPT-2/3 days, where prompting was like 100x harder. We have it so easy now.
-
Bixtral deployment errors on A100 GPUs HuggingFace endpoints
By
–
Hey @michelleyhbn @LucSGeorges just wanted to share that there seem to be errors when using 4x A100 80gbs to host Bixtral on HF endpoints… seems like more than enough RAM, but still crashed — tried gptq etc. No dice, all failed. Not a huge deal for me, but wanted to make sure
