피자먹다잠든무지님이 보내주신 Beyond Classification: Financial Reasoning in State-of-the-Art Language Models을 anysummary.app 에서 우선 가사로 만들었습니다.
그리고 chatgpt로 코드를 만든 후 tuna.voicemod.net 에서 목소리를 입혔습니다.
(아쉽게도 여기서는 코드 조절은 되지 않아 그냥 정해진 음에 맞춰 부른 것입니다.)
Stay-With-Me-with-Voicemod-Text-to-Song.mp4
1절
[Em] In the world of finance, [Am] language models rise,
[D] Their ability to generate, [B7] a pleasant surprise.
[Am] But the field lacks investigation, [Em] datasets are few,
[C] Leaving financial reasoning, [B7] unexplored, it's true.
[Em] Introducing FIOG, a [Am] task that seeks to train,
[D] Language models for investment, [B7] opinions to gain.
[Am] Insights into their efficacy, [Em] their reasoning skill,
[C] And their role in investment, [B7] a thrilling thrill.
[Em] Language models have shown, [Am] reasoning capabilities,
[D] But can they solve financial problems? [B7] Unknown possibilities.
[Am] With parametric or injected knowledge, [Em] they generate views,
[C] But can they scale? That's the [B7] question, no clues.
[Em] Previous research in finance, [Am] focused on classification,
[D] Token or sequence tasks, [B7] lacking exploration.
[Am] But this study aims to dive, [Em] deep to understand,
[C] How language models in finance, [B7] can truly expand.
나머지 가사들
Instruction tuning and synthetic data generation,
Prompting methods, evaluation capability investigation,
All part of the research to uncover,
The potential of language models, like no other.
Incorporating investment information, they reason,
Not just as knowledge databases, but engines for decision-seasoned,
But the lack of suitable datasets poses a challenge,
To develop specialized language models, it's a balance.
The field of finance holds complex nomenclature and steps,
Deviation from general patterns, higher requirements begets,
To train language models for tasks like opinion generation,
Data limitations hinder, keeping it under foundation.
But this research presents a comprehensive investigation,
Into applying language models in finance, a revelation,
With in-context question answering, a novel prompting approach,
Controlling the generation of context, to encroach.
The results reveal the emergence of coherent reasoning,
At 6B parameters, the models start their seasoning,
Improving with better instruction-tuning and more data,
Their financial reasoning capabilities getting even greater.
Lexical and syntactic diversity metrics are used,
To measure richness and coherence, to choose,
The best approaches for generating investment opinions,
Revealing the strengths of language models, their dominion.
The research provides insights into their efficacy,
In the financial domain, engaging in reasoning so craftily,
And while datasets and models still have limitations,
The potential for language models in finance, amazing revelations.
So sing a song of financial reasoning and language models,
Exploring the possibilities, breaking new thresholds,
With FIOG and synthetic data, we advance,
Our understanding of finance, in a captivating trance.