👋 there! My name’s Daniel. A brief blurb about me: I’m a data scientist building GenAI products in the b2b software space. I get excited about technology that transforms how work gets done for the better.
To get in touch, drop me a line at dryanfurman@gmail.com.
Marin Headlands (Oct 2025)
Palace of Fine Arts (Oct 2025)
Golden Gate (Jan 2025)
Ocean Beach (Feb 2022)
Chestnut Street (Feb 2025)
Marin Headlands (Jul 2025)
Tank Hill (Jun 2020)
Big Sur (Dec 2016)
Outskirts of Seattle (Mar 2025)
Marina District (May 2025)
Mallorca Cap de Formentor (Jul 2025)
Mendocino (Apr 2025)
Marin Headlands (Oct 2025)
Palace of Fine Arts (Oct 2025)
Golden Gate (Jan 2025)
Ocean Beach (Feb 2022)
Chestnut Street (Feb 2025)
Marin Headlands (Jul 2025)
Tank Hill (Jun 2020)
Big Sur (Dec 2016)
Outskirts of Seattle (Mar 2025)
Marina District (May 2025)
Mallorca Cap de Formentor (Jul 2025)
Mendocino (Apr 2025)
Marin Headlands (Oct 2025)
Palace of Fine Arts (Oct 2025)
Golden Gate (Jan 2025)
Ocean Beach (Feb 2022)
Chestnut Street (Feb 2025)
Marin Headlands (Jul 2025)
Tank Hill (Jun 2020)
Big Sur (Dec 2016)
Outskirts of Seattle (Mar 2025)
Marina District (May 2025)
Mallorca Cap de Formentor (Jul 2025)
Mendocino (Apr 2025)
LLMs are typically finetuned on chat-templated datasets, which convert conversations into strings ready for text generation. Matching this format is standard at prediction time, as deviations are often noted to hurt performance. Here, we test this theory using an instruction-following benchmark and the best open-source LLMs available at this time.
