This is huge news. Mistral 7B was already the best model in its size class, and these improvements are a huge step up. I’ll be re-training many of my current fine-tunes over this model ASAP.
@mattshumer_
-
Better LLM Reward Model Performance Without Numerical Scales
By
–
If you’re using out of the box LLMs as a reward model/sample rater, here’s a trick to get way better performance: Don’t ask the LLM to rate examples on a numerical scale (i.e. 1-5). The model will almost always choose 1 or 5. Instead, use words as rating options (“very bad”,
-
Axolotl and Together Compute for Model Fine-tuning
By
–
I use Axolotl, but services like @togethercompute might be easier. Last I checked, they don’t support only training on outputs, which is important. May have changed since!
-
Fine-tuning LLMs: Progressive Data Selection Strategy
By
–
High alpha technique for fine-tuning LLMs: First, fine-tune your model on all your data Then, use AI to pick a smaller set of really good examples, and tune on them Finally, hand-pick and tune on a few hundred incredible examples >> Huge performance gains <<
-
Claude-Investor Meme Coin Version Announced
By
–
Fine since you all are asking I’ll make a meme coin version of Claude-Investor
-
HyperWrite: AI Writing Assistant with Fast Internet Access
By
–
Use HyperWrite, it’s a really great assistant (especially for writing) + has super-fast internet access
-
Large Models Improve Smaller Model Performance via Prompt Engineering
By
–
Rohit just merged this to the `gpt-prompt-engineer` repo. Now, you can use any large model to improve the performance of a smaller model! https://
github.com/mshumer/gpt-pr
ompt-engineer
… -
RLHF Trade-offs: Control vs Capability in Advanced LLMs
By
–
Many possible reasons. I’d speculate that they want something more controlled because they have a lot to lose being in the #1 spot. RLHF can help capabilities if done right, but GPT-4 is just way overdone.
-
HyperWrite vs Claude 3 Opus: Task-Specific AI Tool Selection
By
–
HyperWrite for writing + research, Claude 3 Opus for code, reasoning tasks (for now)
-
Claude 3 Opus vs GPT-4: Model Capabilities Beyond Assistant Performance
By
–
Claude 3 is a much better assistant than GPT-4. But it’s NOT a much better model, as most people seem to be thinking. The production version of GPT-4 has been RLHF-ed to hell, but the pre-release alpha was far more powerful. It felt very, very close to how Claude 3 Opus feels