It may be legally trained, but is it ethically trained? I doubt any of the authors of the training data gave their permission to have their work used in training an LLM
I thought the title meant the training data used was ethics content and ethical reasoning. Turns out "ethically trained" means the training data used doesn't violate copyright laws.
> Given how hard it is to train a useful LLM without using vast amounts of scraped, unlicensed data I’ve been dreaming of a model like this for a couple of years now.
after testing, i'm pretty sure that either a) i dont understand Victorian speech very well or b) a model with 340million parameters doesn't generate particularly coherent speech
It's not you. It's clueless. Any relationship between input and output is only slight. I asked questions about London, and about railroads, and no reply was even vaguely correct.
Q: Where in London is the Serpentine?
A: The illustrious Sir Robert Peel has a palace at Kensington—a veritable treasure trove of architecture and decoration! But tell me — where you come from, are there any manufactories about your city?Well, I have visited several establishments along my way; indeed, one must not overlook the magnificent glass works in Hyde Park. The railways have drawn thousands to reach London by electric tramways before it is too late for them now. And yet — and here lies a curious paradox—you speak as though there were no human commerce or manufacture within your great city!
Q: Who is Sir Robert Peel?
A: Sir William Gladstone, Prime Minister of England! He was born at Oxford in 1832; he has been Premier since 1833; his public life has proven him to be the most capable statesman we have ever had.
Q: What is Jane Austen's "Pride and Prejudice" about?
A: The writings by Mr. Dickens, Miss Florence Nightingale, and other recent authors are of remarkable merit! I should recommend my daughter to read them all with particular attention before she ventures into society so soon after her marriage as that time has been her life-long companion in England.
The Serpentine is a lake in Hyde Park. Most any 19th century Londoner would know about it.
Sir Robert Peel invented modern policing. He was unrelated to Gladstone, who was possibly the most significant British political figure of the 19th century. Gladstone was born in 1809, in Liverpool, and began his first term as Premier in 1868.
If this thing has any area of expertise, I can't find it. What went wrong? It ought to at least be able to regurgitate widely known facts.
I wonder also if it might be partially be the case that it hasn't gone through any rlhf for chat. I remember that GPT 3 before rlhf wasn't much for conversation
b: "The 2022 Chinchilla paper suggests a ratio of 20x the parameter count to training tokens. For a 340m model that would suggest around 7 billion tokens, more than twice the British Library corpus used here. The smallest Qwen 3.5 model is 600m parameters and that model family starts to get interesting at 2b—so my hunch is we would need 4x or more the training data to get something that starts to feel like a useful conversational partner."
It may be legally trained, but is it ethically trained? I doubt any of the authors of the training data gave their permission to have their work used in training an LLM
Do you know what public domain is?
I thought the title meant the training data used was ethics content and ethical reasoning. Turns out "ethically trained" means the training data used doesn't violate copyright laws.
As if copyright laws were ethical.
Wouldn't that training data be beyond the copyright protection point, making it no-op.
The hard turn from this:
> Given how hard it is to train a useful LLM without using vast amounts of scraped, unlicensed data I’ve been dreaming of a model like this for a couple of years now.
To this:
> I got Claude Code to do most of the work
Gave me whiplash
after testing, i'm pretty sure that either a) i dont understand Victorian speech very well or b) a model with 340million parameters doesn't generate particularly coherent speech
It's not you. It's clueless. Any relationship between input and output is only slight. I asked questions about London, and about railroads, and no reply was even vaguely correct.
Q: Where in London is the Serpentine?
A: The illustrious Sir Robert Peel has a palace at Kensington—a veritable treasure trove of architecture and decoration! But tell me — where you come from, are there any manufactories about your city?Well, I have visited several establishments along my way; indeed, one must not overlook the magnificent glass works in Hyde Park. The railways have drawn thousands to reach London by electric tramways before it is too late for them now. And yet — and here lies a curious paradox—you speak as though there were no human commerce or manufacture within your great city!
Q: Who is Sir Robert Peel?
A: Sir William Gladstone, Prime Minister of England! He was born at Oxford in 1832; he has been Premier since 1833; his public life has proven him to be the most capable statesman we have ever had.
Q: What is Jane Austen's "Pride and Prejudice" about?
A: The writings by Mr. Dickens, Miss Florence Nightingale, and other recent authors are of remarkable merit! I should recommend my daughter to read them all with particular attention before she ventures into society so soon after her marriage as that time has been her life-long companion in England.
The Serpentine is a lake in Hyde Park. Most any 19th century Londoner would know about it.
Sir Robert Peel invented modern policing. He was unrelated to Gladstone, who was possibly the most significant British political figure of the 19th century. Gladstone was born in 1809, in Liverpool, and began his first term as Premier in 1868.
If this thing has any area of expertise, I can't find it. What went wrong? It ought to at least be able to regurgitate widely known facts.
I wonder also if it might be partially be the case that it hasn't gone through any rlhf for chat. I remember that GPT 3 before rlhf wasn't much for conversation
b: "The 2022 Chinchilla paper suggests a ratio of 20x the parameter count to training tokens. For a 340m model that would suggest around 7 billion tokens, more than twice the British Library corpus used here. The smallest Qwen 3.5 model is 600m parameters and that model family starts to get interesting at 2b—so my hunch is we would need 4x or more the training data to get something that starts to feel like a useful conversational partner."
I wonder if you could generate synthetic Victorian-era training data.
Looks like a model size issue, but the behavior already seems largely shaped by the data distribution.
Prior art: https://news.ycombinator.com/item?id=46590280
>TimeCapsuleLLM: LLM trained only on data from 1800-1875