Jørgen Lund
@jaalu.bsky.social
Industry Ph.D. student in ML, DIPS AS, UiT The Arctic University of Norway
github.com/jaalu | he/him
github.com/jaalu | he/him
At 56M parameters, naive quantization *may* not be the best way to go
November 13, 2025 at 3:59 PM
At 56M parameters, naive quantization *may* not be the best way to go
@ai2.bsky.social's OLMoTrace (arxiv.org/pdf/2504.07096) is the more serious approach to this, they match spans from the output to spans in the training corpora, and the paper does find links between answers and specific sources, but as far as I can tell they do not intervene in the generation itself
![A picture of OLMoTrace showing the OLMo 2 model answering the question "How does a CPU work?", with the phrase "[the program counter] holds the location of the next instruction to be executed" highlighted and linked to a document in the training corpus](https://cdn.bsky.app/img/feed_thumbnail/plain/did:plc:racxi7k4vbwf37jqgus3l4u7/bafkreiczbpikttjsf3sxgta7742vct4xmg7yur5ejvp2zbdzl7hgpmweby@jpeg)
October 2, 2025 at 8:17 AM
@ai2.bsky.social's OLMoTrace (arxiv.org/pdf/2504.07096) is the more serious approach to this, they match spans from the output to spans in the training corpora, and the paper does find links between answers and specific sources, but as far as I can tell they do not intervene in the generation itself
In the specific project the dataset seems to be books, legal documents and contemporary text sourced from Project Gutenberg
There is a surprising amount of text available though, Harvard's Institutional Books dataset - arxiv.org/pdf/2506.08300 - has >470K texts dating to the 1800s
There is a surprising amount of text available though, Harvard's Institutional Books dataset - arxiv.org/pdf/2506.08300 - has >470K texts dating to the 1800s
August 26, 2025 at 11:47 AM
In the specific project the dataset seems to be books, legal documents and contemporary text sourced from Project Gutenberg
There is a surprising amount of text available though, Harvard's Institutional Books dataset - arxiv.org/pdf/2506.08300 - has >470K texts dating to the 1800s
There is a surprising amount of text available though, Harvard's Institutional Books dataset - arxiv.org/pdf/2506.08300 - has >470K texts dating to the 1800s