Is Fine-Tuning Still Needed? LLMs, RAG, & LoRA

Comments · 165

• 遊蕩

  @遊蕩者 · 2 months ago

  distillation-oriented finetuning may be the most popular trend in community, and Chinese AI labs.<br>As long as tasks are simple enough, we can use smart small model to replace large frontier one.

  11

• TT

  @ttt6262 · 2 months ago

  I definitely think fine-tuning is still very relevant, especially for smaller models with specialized tasks. Of course, the very large models are significantly better in every respect than these smaller models, which is precisely why fine-tuning is so important for smaller models—to ensure they perform their task at least as well as the very large models, while running locally and offering significantly better data protection than the large cloud models.

  27

• JE

  @jeffreywhewhetu · 2 months ago

  With what I just learnt from the video? I think privacy would be one of the major reasons to do fine tuning if you ask me<br>Thank you for updating us on the state of AI via your videos, IBM!!

  42

• CH

  @chrismoore4803 · 2 months ago

  100% agree with what we’ve seen in the field. The one Case for fine-tuning that I have also seen do well is capturing tone which is used in a specific industry. Your legal example may also be suggestive of this phenomenon. You mentioned that lawyers liked the output of the fine-tune model. Clearly that doesn’t necessarily make the model quantitatively better or worse, but we have seen that fine-tuning captures. The tone of voice used in a field, and it seems to resonate with experts in a way that prompt guidance struggles to capture.

  10

• EQ

  @equan18 · 2 months ago

  These videos are fantastic, and the best education series on AI! Keep these coming!

  23

• RE

  @RealMcDudu · 2 weeks ago

  The video is probably referring to Harvey’s legal AI system. Newer general-purpose frontier models eventually outperformed Harvey’s original specialized model - but that’s not the same as showing that general models outperform a newly fine-tuned model based on the same generation of technology. Specialization may still provide an advantage; the problem is that frontier models are improving so quickly that this advantage can close within a year or even less.

  2

• SE

  @SergeZIEHI · 1 month ago

  So mindful &amp; precious time grasping this unvaluable content, delighting every second... One of the best if not the best Nerd content out there.

  1

• FE

  @felix5310 · 2 months ago

  I found finetuning to also be very useful if you want to force a certain writing style. Giving examples also works, but if you want it to become really consistent, nothing beats a good finetuning run.

  2

• TE

  @Tenebrisuk · 2 months ago

  Fine tuning still has a place for other tasks such as classification and sentiment analysis, where you are likely going to have a more consistent and far more efficient pipeline, not to mention the impacts in terms of compute (which relate to both financial and environmental cost). &nbsp;Not every task is best suited with a generative decoder model. &nbsp;Even the retrieval augmented generation, the retrieval part can occasionally benefit from finetuning.

  45

• GO

  @gorairakesh · 2 months ago (edited)

  Also, sometimes a fresh mind helps in problem solving. &nbsp;This made me remember an incident where an IT architect and a team of specialists &nbsp;couldn&apos;t resolve an issue and finally they had to involve an external consultant who didn&apos;t have any prior knowledge of that specific project. But he was able to resolve the issue on account of his experience in other similar projects. So, in this context I think the base Frontier models bring freshness of mind, who can think deep and independently.

  1

• 36

  @360captureit · 2 months ago

  As always, Martin thought-provoking - thank you

• NI

  @NikhilGoyal-k9m · 3 weeks ago

  exceptional! thank you!