Depends on the LLM and its training data. Will have to think more about the second question.
@mattshumer_
-
Fine-tune Mistral-7B with synthetic prompt-response datasets
By
–
Switch all prompts to mistral-large. Run the repo many times, collect the prompts / responses for each input / output. Format as JSONL, train a Mistral-7B or similar on all of them, and slot it in.
-
API Optimization and qLoRA Fine-tuning Techniques
By
–
On APIs, it’s hard but if you’re going after a specific use-case it just takes a bit of tinkering with the prompt. If not, just do a qLoRA with a couple hundred negative examples and you’re all set.
-
Prompt First: When to Fine-tune AI Models
By
–
Always, always try prompting first. The only time you should fine-tune is when you NEED to.
-
High-Quality Data Training for AI Models
By
–
Really good data with that style. You don’t need a ton of it (few hundred examples is totally fine), just make sure it’s really high-quality.
-
HyperWrite Internal Model A/B Testing in Production
By
–
Yes! One of our most recent internal HyperWrite models is trained for this. We’re currently A/B testing this in production, so if you’re lucky, you’ll get this! cc @JasonKuperberg more validation we should improve this 🙂
-
LLM API Strategy: Wait for PMF Before Fine-tuning
By
–
If the LLM API can do your task, if you can afford the costs, wait to fine-tune until you have PMF and need to optimize.
-
Chaining Smaller Prompts vs Mega Prompts Strategy
By
–
Usually stringing smaller prompt outputs together with code is better, but mega prompts are a close and easier second.
-
Fine-tuning AI models in 10 minutes with Runpod
By
–
10 minutes? Pretty doable actually, if the data is ready. Spin up a Runpod machine. While that’s starting, upload dataset to huggingface. Then write a yaml for the run config (Axolotl) and upload to Gist. By that time the machine is ready. Run this command: pip uninstall -y