Theres lots more to do to understand CFT better, and build on it to create better post-training methods to fine-tune large language models.

Reach out to me or Ethan if you're interested in collaborating on this or pushing this idea to new domains and problems!

📖 We’ve also open-sourced OpenMedText, integrating 121K biomedical articles & 29 medical textbooks to push future research in domain-adaptive fine-tuning in biomedicine.

🔧 We "negative" and "adaptive" prompts, confirming that the semantic content of prompts changes and impacts fine-tuning effectiveness.

📊 Results: On medical benchmarks, CFT improves accuracy by ~2.25% over CPT; in finance, it boosts performance by ~4.32%! Importantly, these gains scale effectively with larger models. 📈

Check out Appendix E.1 for preliminary results on GEMINI Flash 1.5M!

🏥 We tested this idea in biomedical (using newly curated OpenMedText dataset of journals & textbooks!) and financial data—CFT significantly outperforms continued pretraining (CPT) and instruction fine-tuning (IFT) in zero-shot settings.

🎓 Instead of using Q&A as in instruction tuning, CFT uses reflective instructions (e.g., "Reflect on how what you will see changes what you know...") motivated by how humans learn.

💡Contextual finetuning (CFT) uses contextual prompts during fine-tuning to adaptively change the semantic understanding that LLMs leverage during the process of learning new information.

🚀 Problem: Language models struggle with rapidly evolving info and context in fields like medicine & finance. We need ways to teach LLMs new information and control how they absorb this knowledge.

🔍 Insight: Why not explain and teach LLMs how to learn?

If it helps, I usually learn something new (either directly or from further digging) about the behavior of markets.

Rocking that @ Gmail address!

I thought about this a bit, I think helping PhD students close the translational gap from research to deployment (in industry or their own startups), particularly if they don't want to go into academia, is one way forward.

Are you around at Neurips? Would love to say hi and catch up!

Finally, if you're interested in understanding how to leverage energy-based normalizing flows, check out Lance's work on Meow (chienfeng-hub.github.io/meow/)

He'll be presenting on Dec. 12, 11:00 AM–2:00 PM at West Ballroom A-D #6403

🧵(7/7)

@nikitadhawan.bsky.social developed NATURAL (www.cs.toronto.edu/~nikita/natu...) with @cottascience.bsky.social , Karen & @cmaddis.bsky.social. Its an end-to-end pipeline that starts from raw-text data and ends with a causal (**) effect associated with an intervention.

(**) conditions apply
🧵(6/7)

b] ~Billions of dollars each year are spent on trials to assess interventions.

Can we use crowdsourced data to know which intervention is likely to work ahead of time?

Doing so requires answering a causal question!

But the data to answer this question is locked in unstructured text.

🧵(5/7)

Find Vahid to learn more about in-context causal inference and lots of other cool problems that he spends his time thinking about!

🧵(4/7)

In arxiv.org/abs/2404.07266, Vahid shows how to use offline expert data with unobserved confounding to guide decision making using a nonparametric prior to guide learning policies for bandits, MDPs, and POMDPs.

Thu 12 Dec 4:30 - 7:30 pm PST 📷 West Ballroom A-D Poster #6708

🧵(3/7)

a] Today, we learn from data and treat it as ground truth -- should we?

A doctor often knows more about their patient than is represented in electronic medical records.

A teacher knows more about their students than what their grades suggest.

🧵(2/7)